* Update API Reference for Pregel
* Add conceptual page for Pregel
* The content for the two is very similar at the moment (i.e.,
duplicated content). This is usually a bad sign, but in this case I'm OK
duplicating information along both paths since the underlying algorithm
sets us apart from other implementations.
This PR removes the unused imports Literal and TypedDict from the typing
module.
These imports were not referenced in the code.
```python
from typing import Literal, TypedDict
```
This PR removes the following changes:
* notebooks that were converted to markdown
* mkdocs.yml file to reference the ipython notebooks rather than the
markdown files
* Makefile install vercel reverted
* hooks for markdown-exec
* notebook conversion jinja2 templates (for converting notebooks to
markdown exec format)
Hi, I am a student and was going through the tutorial. While trying to
understand the different components by reading the docstring found this
super minor typo 😄 . I hope to contribute more meaningful changes in
future 😸
Currently when using RemoteGraph the recursion_limit cannot be set, due
to the sanitize_config.
---------
Co-authored-by: Simon Moxon <simon@together.ly>
Co-authored-by: Vadym Barda <vadim.barda@gmail.com>
Currently a ChatPromptTemplate cannot be used as a `prompt` for
`create_react_agent` without complaints from type checkers, although it
is supported by `model` as input.
Add the missing types to remove the warning.
---------
Co-authored-by: vbarda <vadym@langchain.dev>
- Now supporting local dependencies in directories that are not
contained in the docker context (ie. outside the folder containing
langgraph.json)
- This is achieved by passing each parent directorty as an additional
context to docker build
- This makes it a lot easier to build projects contained in monorepos
where you need to include some sibling/parent folder as a dependency
- Also include additional comments in the generated dockerfile to
delimit each section
- Now supporting local dependencies in directories that are not contained in the docker context (ie. outside the folder containing langgraph.json)
- This is achieved by passing each parent directorty as an additional context to docker build
- This makes it a lot easier to build projects contained in monorepos where you need to include some sibling/parent folder as a dependency
- Also include additional comments in the generated dockerfile to delimit each section
* Add ast parsing to determine whether we should include result="ansi".
It's not meant to be perfect, but will hopefully catch the most common
cases. Still requires manual review.
* Ideally we could suppress output in markdown-exec in the future.
* Adds another notebook conversion
* Fix up some edge cases for handling links in notebooks. Notebooks
links were using a different convention than markdown links.
We'll need to push additional logic to use an appropriate suffix (.md or
.ipynb) for cross-references between how-to guides (though these should
be rare).
* Add testing step to to docs build pipeline
* Requires updating import structure in some place
* Add simple unit test to cover some logic with highlights
* Adds a conversion script from ipython notebook to markdown.
* Replaces one ipython notebook (create react agent) with a markdown file for testing.
---------
Co-authored-by: Ben Burns <803016+benjamincburns@users.noreply.github.com>
Adding open source web researcher agent to the third party page
---------
Co-authored-by: Eugene Yurtsev <eyurtsev@gmail.com>
Co-authored-by: Eugene Yurtsev <eugene@langchain.dev>
StateType, UpdateType is set on the Client rather than on
`RunsClient.stream` because of lack of partial type arguments
application.
This PR also describes the message serialization format emitted by
LangGraph Server (which may change ie. converting `type` to `role`).
Avoiding direct import of `@langchain/core` for the core LangGraph SDK
client, thus these types were copied from `@langchain/core` (a script is
used to aid with keeping track with core)
(Keep MemorySaver around for backwards compatibility)
"MemorySaver" is ambiguous: is it saving memories? Where is it saving
memories to?
InMemorySaver aligns naming InMemoryStore as well as similar LangChain
objects (InMemoryVectorStore, etc.)
Was trying to learn the Multi Agent Workflow examples and encountered
some errors, which I fixed by editing these:
* Added missing state for Team1, and importing `Command`
* `ValueError: Node `LangGraph` already present.`: Seems to happen we
add the `team_1_graph` node without giving it a name, it will default to
the name `LangGraph`. Solved by giving the sub-graph a name when
building the top-level supervisor.
* Added the edges for the graph to feedback to the top level supervisor
to decide whether it still needs to relegate the task to other nodes or
end from there
---------
Co-authored-by: Vadym Barda <vadim.barda@gmail.com>
Alternative to https://github.com/langchain-ai/langgraph/pull/3124
Currently if a tool interrupts, the entire tool node executes again
after resuming. So tools can get executed twice if parallel tool calls
are generated. Here we allow ToolNode to accept tool calls, so we can
use the `Send` API to distribute the tool calls to multiple instances of
the tool node.
```python
from langchain_anthropic import ChatAnthropic
from langchain_core.tools import tool
from langgraph.checkpoint.memory import MemorySaver
from langgraph.prebuilt import create_react_agent
from langgraph.types import Command, Send, interrupt
@tool
def human_assistance(query: str) -> str:
"""Request assistance from a human."""
human_response = interrupt({"query": query})
return human_response["data"]
@tool
def get_weather(location: str) -> str:
"""Use this tool to get the weather."""
return "It's sunny!"
tools = [get_weather, human_assistance]
llm = ChatAnthropic(model="claude-3-5-sonnet-20240620")
agent = create_react_agent(
llm,
tools,
checkpointer=MemorySaver(),
tool_call_parallelism="parallel_tool_nodes",
)
user_input = (
"Could you please (1) request assistance for building an AI agent "
"from a human, and (2) search for the weather in Boston, MA? "
"Generate two tool calls at once."
)
config = {"configurable": {"thread_id": "1"}}
for event in agent.stream(
{"messages": [{"role": "user", "content": user_input}]},
config,
stream_mode="values",
):
event["messages"][-1].pretty_print()
```
```
...
```
```python
human_response = "You should check out LangGraph to build your agent."
human_command = Command(resume={"data": human_response})
for event in agent.stream(human_command, config, stream_mode="values"):
event["messages"][-1].pretty_print()
```
---------
Co-authored-by: Vadym Barda <vadym@langchain.dev>
- The result of these doesnt change once a node is created, and it's
fairly expensive to run, so great thing to cache
- There's a variety of errors that can come from inspecting the source
code of a function (part of what this does) so adding a catch-all
try-except block as this should be best-effort, not crash your graph
- The result of these doesnt change once a node is created, and it's fairly expensive to run, so great thing to cache
- There's a variety of errors that can come from inspecting the source code of a function (part of what this does) so adding a catch-all try-except block as this should be best-effort, not crash your graph
* Concepts page for the functional API
* How-to guides that show functional API implementations
* API reference for entrypoint, task, entrypoint.final
* Add functional API version to the workflows
---------
Co-authored-by: Vadym Barda <vadym@langchain.dev>
Co-authored-by: ccurme <chester.curme@gmail.com>
When I cloned langgraph example, I was not able to run the code because
of the requirements.txt file. The path was set to posix expression which
was not working in windows OS.
I have updated the path to posix expression so that it can work in
windows OS as well.
Try running `langgraph-example` in windows using the langgraph-cli in
windows OS. It was working for linux not in windows.
- async tests are placed in test_pregel_async, not in test_pregel
- to avoid tests placed in wrong file being accidentally skipped i've
added the auto-async mark to sync test file
- this was not possible in async where all done callbacks are called in
next tick
- in sync case this would manifest as the first task done callback
seeing counter == 1 and thus setting event
- the fix is to unset the event whenever a task is scheduled
- When using an async entrypoint you can now freely mix and match sync
and async tasks with a uniform api (ie all tasks return a sync or async
future depending on context)
- Fix issues with scheduling deeply nested tasks (use threadsafe methods
to schedule coroutines and create futures)
So that you can call agent.nodes['agent'].invoke({'messages': []})
without needing to specify is_last_step. very helpful for evaluating
just the model node of the agent
- async tests are placed in test_pregel_async, not in test_pregel
- to avoid tests placed in wrong file being accidentally skipped i've added the auto-async mark to sync test file
- this was not possible in async where all done callbacks are called in next tick
- in sync case this would manifest as the first task done callback seeing counter == 1 and thus setting event
- the fix is to unset the event whenever a task is scheduled
- When using an async entrypoint you can now freely mix and match sync and async tasks with a uniform api (ie all tasks return a sync or async future depending on context)
- Fix issues with scheduling deeply nested tasks (use threadsafe methods to schedule coroutines and create futures)
- both issues are related to the fact that waiters for futures are
notified of completion before "done" callbacks are called
- 1st issue manifested as interrupt stream event being emitted before
the result of a task that logically finished first (it's in the line
above in body of the entrypoint function) -> this is solved by always
returning to use code a fresh future chained on the original future,
because chaining is done via done callbacks (therefore the chained
future will only resolve after done callbacks of the original feature
are called)
- 2nd issue mainfested as sometimes (very rarely) the last stream event
not being printed before stream() finishes. this is solved by ensuring
we only return out of PregelRunner.tick() once all "done" callbacks are
called, previously we were approximating this through use of
asyncio.sleep(0) / time.sleep(0). The new solution instead waits on a
threading/asyncio.Event which will only be set by the last "done"
callback to fire
- this PR also disables incomplete support for calling sync tasks from
async entrypoints
- both issues are related to the fact that waiters for futures are notified of completion before "done" callbacks are called
- 1st issue manifested as interrupt stream event being emitted before the result of a task that logically finished first (it's in the line above in body of the entrypoint function) -> this is solved by always returning to use code a fresh future chained on the original future, because chaining is done via done callbacks (therefore the chained future will only resolve after done callbacks of the original feature are called)
- 2nd issue mainfested as sometimes (very rarely) the last stream event not being printed before stream() finishes. this is solved by ensuring we only return out of PregelRunner.tick() once all "done" callbacks are called, previously we were approximating this through use of asyncio.sleep(0) / time.sleep(0). The new solution instead waits on a threading/asyncio.Event which will only be set by the last "done" callback to fire
1. The inputs into foo do not affect any state behavior
2. `previous` always reflects the previous return value from the
function
3. Anything can be returned and that will be the new state for the
function on the next iteration
4. This API is not meant to support reducers in the inputs/state
```python
from langgraph.func import entrypoint
states = []
# In this version reducers do not work
@entrypoint(checkpointer=MemorySaver())
def foo(inputs, *, previous: Any) -> Any:
states.append(previous)
return {"previous": previous, "current": inputs}
config = {"configurable": {"thread_id": "1"}}
foo.invoke({"a": "1"}, config)
foo.invoke({"a": "2"}, config)
foo.invoke({"a": "3"}, config)
assert states == [
None,
{"current": {"a": "1"}, "previous": None},
{"current": {"a": "2"}, "previous": {"current": {"a": "1"}, "previous": None}},
]
```
Currently, if you're viewing a how-to guide and you click "How-to
Guides" in the sidebar, you aren't navigated back to the index page (it
will work if you click on a different guides section). To get back to
the index page, you need to scroll up and click the breadcrumbs.
After this change, clicking the link in the sidebar should navigate you
to the index page regardless of the page you are viewing.
Only side-effect from what I can tell is that "Home > Introduction" just
becomes "**Home**", which I think is fine (maybe preferable).
Before:
After:
- order was incorrectly based on task id, instead of the correct task
path
- this requires storing task paths on checkpointers
- addition of task_path to put_writes is made backwards compatible by
checking signature on call, and treating it as an optional arg
- order was incorrectly based on task id, instead of the correct task path
- this requires storing task paths on checkpointers
- addition of task_path to put_writes is made backwards compatible by checking signature on call, and treating it as an optinal arg
Some docs layout improvements to help guide user journey.
Currently we have `Home | Tutorials | How-tos | Concepts | Reference` in
top-level horizontal navigation bar.
Here we make these updates:
- Top-level horizontal navigation bar is just `Home | API Reference`
- Add vertical sidebar to `Home` with sections:
- Introduction
- Get started
- Guides
- Resources
`Get Started` contains quickstarts for LG and LG Platform / deployment.
These are tutorials in Diataxis terms.
`Guides` contains index pages for how-tos, concepts, tutorials.
Advantage of this organization is that users are directed naturally down
the sidebar from Intro -> Get started -> How-tos, which is roughly how
we expect them to proceed.
This also makes deployment info more accessible as it is highlighted in
the "Getting started" section.
When I tried to follow the How-to guide for [How to add semantic search
to your agent's
memory](https://langchain-ai.github.io/langgraph/how-tos/memory/semantic-search/#using-in-create_react_agent)
using `create_react_agent`, I got this error message when my agent used
the tool:
```python
1 validation error for upsert_memory
store
Field required [type=missing, input_value={'content': '@jimmy works...ny.', 'memory_id': None}, input_type=dict]
For further information visit https://errors.pydantic.dev/2.10/v/missingTraceback (most recent call last):
File "/usr/local/lib/python3.9/site-packages/langchain_core/tools/base.py", line 688, in run
tool_args, tool_kwargs = self._to_args_and_kwargs(tool_input, tool_call_id)
File "/usr/local/lib/python3.9/site-packages/langchain_core/tools/base.py", line 611, in _to_args_and_kwargs
tool_input = self._parse_input(tool_input, tool_call_id)
File "/usr/local/lib/python3.9/site-packages/langchain_core/tools/base.py", line 532, in _parse_input
result = input_args.model_validate(tool_input)
File "/usr/local/lib/python3.9/site-packages/pydantic/main.py", line 627, in model_validate
return cls.__pydantic_validator__.validate_python(
pydantic_core._pydantic_core.ValidationError: 1 validation error for upsert_memory
store
Field required [type=missing, input_value={'content': '@jimmy works...ny.', 'memory_id': None}, input_type=dict]
For further information visit https://errors.pydantic.dev/2.10/v/missing
```
I believe it’s because the graph did not inject the store into the tool
if we use `InjectedToolArg`.
When looking at the guide for [How to pass runtime values to
tools](https://langchain-ai.github.io/langgraph/how-tos/pass-run-time-values-to-tools/),
it suggests to use `InjectedStore` with `create_react_agent`. After
changing my code to use `InjectedStore`, my agent was able to save to
the store.
```python
class WeatherResponse(BaseModel):
"""Respond to the user with this"""
temperature: float = Field(description="The temperature in fahrenheit")
wind_direction: str = Field(
description="The direction of the wind in abbreviated form"
)
wind_speed: float = Field(description="The speed of the wind in mph")
@tool
def get_weather(city: Literal["nyc", "sf"]):
"""Use this to get weather information."""
if city == "nyc":
return "It is cloudy in NYC, with 5 mph winds in the North-East direction and a temperature of 70 degrees"
elif city == "sf":
return "It is 75 degrees and sunny in SF, with 3 mph winds in the South-East direction"
else:
raise AssertionError("Unknown city")
model = ChatOpenAI()
tools = [get_weather]
agent_with_structured_output = create_react_agent(model, tools, response_format=WeatherResponse)
agent_with_structured_output.invoke({"messages": [("user", "what's the weather in nyc?")]})
```
```pycon
{
'messages': [...],
'structured_response': WeatherResponse(temperature=70.0, wind_directon='NE', wind_speed=5.0)
}
```
Bumps [jinja2](https://github.com/pallets/jinja) from 3.1.4 to 3.1.5.
<details>
<summary>Release notes</summary>
<p><em>Sourced from <a
href="https://github.com/pallets/jinja/releases">jinja2's
releases</a>.</em></p>
<blockquote>
<h2>3.1.5</h2>
<p>This is the Jinja 3.1.5 security fix release, which fixes security
issues and bugs but does not otherwise change behavior and should not
result in breaking changes compared to the latest feature release.</p>
<p>PyPI: <a
href="https://pypi.org/project/Jinja2/3.1.5/">https://pypi.org/project/Jinja2/3.1.5/</a>
Changes: <a
href="https://jinja.palletsprojects.com/changes/#version-3-1-5">https://jinja.palletsprojects.com/changes/#version-3-1-5</a>
Milestone: <a
href="https://github.com/pallets/jinja/milestone/16?closed=1">https://github.com/pallets/jinja/milestone/16?closed=1</a></p>
<ul>
<li>The sandboxed environment handles indirect calls to
<code>str.format</code>, such as by passing a stored reference to a
filter that calls its argument. <a
href="https://github.com/pallets/jinja/security/advisories/GHSA-q2x7-8rv6-6q7h">GHSA-q2x7-8rv6-6q7h</a></li>
<li>Escape template name before formatting it into error messages, to
avoid issues with names that contain f-string syntax. <a
href="https://redirect.github.com/pallets/jinja/issues/1792">#1792</a>,
<a
href="https://github.com/pallets/jinja/security/advisories/GHSA-gmj6-6f8f-6699">GHSA-gmj6-6f8f-6699</a></li>
<li>Sandbox does not allow <code>clear</code> and <code>pop</code> on
known mutable sequence types. <a
href="https://redirect.github.com/pallets/jinja/issues/2032">#2032</a></li>
<li>Calling sync <code>render</code> for an async template uses
<code>asyncio.run</code>. <a
href="https://redirect.github.com/pallets/jinja/issues/1952">#1952</a></li>
<li>Avoid unclosed <code>auto_aiter</code> warnings. <a
href="https://redirect.github.com/pallets/jinja/issues/1960">#1960</a></li>
<li>Return an <code>aclose</code>-able <code>AsyncGenerator</code> from
<code>Template.generate_async</code>. <a
href="https://redirect.github.com/pallets/jinja/issues/1960">#1960</a></li>
<li>Avoid leaving <code>root_render_func()</code> unclosed in
<code>Template.generate_async</code>. <a
href="https://redirect.github.com/pallets/jinja/issues/1960">#1960</a></li>
<li>Avoid leaving async generators unclosed in blocks, includes and
extends. <a
href="https://redirect.github.com/pallets/jinja/issues/1960">#1960</a></li>
<li>The runtime uses the correct <code>concat</code> function for the
current environment when calling block references. <a
href="https://redirect.github.com/pallets/jinja/issues/1701">#1701</a></li>
<li>Make <code>|unique</code> async-aware, allowing it to be used after
another async-aware filter. <a
href="https://redirect.github.com/pallets/jinja/issues/1781">#1781</a></li>
<li><code>|int</code> filter handles <code>OverflowError</code> from
scientific notation. <a
href="https://redirect.github.com/pallets/jinja/issues/1921">#1921</a></li>
<li>Make compiling deterministic for tuple unpacking in a <code>{% set
... %}</code> call. <a
href="https://redirect.github.com/pallets/jinja/issues/2021">#2021</a></li>
<li>Fix dunder protocol (<code>copy</code>/<code>pickle</code>/etc)
interaction with <code>Undefined</code> objects. <a
href="https://redirect.github.com/pallets/jinja/issues/2025">#2025</a></li>
<li>Fix <code>copy</code>/<code>pickle</code> support for the internal
<code>missing</code> object. <a
href="https://redirect.github.com/pallets/jinja/issues/2027">#2027</a></li>
<li><code>Environment.overlay(enable_async)</code> is applied correctly.
<a
href="https://redirect.github.com/pallets/jinja/issues/2061">#2061</a></li>
<li>The error message from <code>FileSystemLoader</code> includes the
paths that were searched. <a
href="https://redirect.github.com/pallets/jinja/issues/1661">#1661</a></li>
<li><code>PackageLoader</code> shows a clearer error message when the
package does not contain the templates directory. <a
href="https://redirect.github.com/pallets/jinja/issues/1705">#1705</a></li>
<li>Improve annotations for methods returning copies. <a
href="https://redirect.github.com/pallets/jinja/issues/1880">#1880</a></li>
<li><code>urlize</code> does not add <code>mailto:</code> to values like
<code>@a@b</code>. <a
href="https://redirect.github.com/pallets/jinja/issues/1870">#1870</a></li>
<li>Tests decorated with <code>@pass_context</code> can be used with the
<code>|select</code> filter. <a
href="https://redirect.github.com/pallets/jinja/issues/1624">#1624</a></li>
<li>Using <code>set</code> for multiple assignment (<code>a, b = 1,
2</code>) does not fail when the target is a namespace attribute. <a
href="https://redirect.github.com/pallets/jinja/issues/1413">#1413</a></li>
<li>Using <code>set</code> in all branches of <code>{% if %}{% elif %}{%
else %}</code> blocks does not cause the variable to be considered
initially undefined. <a
href="https://redirect.github.com/pallets/jinja/issues/1253">#1253</a></li>
</ul>
</blockquote>
</details>
<details>
<summary>Changelog</summary>
<p><em>Sourced from <a
href="https://github.com/pallets/jinja/blob/main/CHANGES.rst">jinja2's
changelog</a>.</em></p>
<blockquote>
<h2>Version 3.1.5</h2>
<p>Released 2024-12-21</p>
<ul>
<li>The sandboxed environment handles indirect calls to
<code>str.format</code>, such as
by passing a stored reference to a filter that calls its argument.
:ghsa:<code>q2x7-8rv6-6q7h</code></li>
<li>Escape template name before formatting it into error messages, to
avoid
issues with names that contain f-string syntax.
:issue:<code>1792</code>, :ghsa:<code>gmj6-6f8f-6699</code></li>
<li>Sandbox does not allow <code>clear</code> and <code>pop</code> on
known mutable sequence
types. :issue:<code>2032</code></li>
<li>Calling sync <code>render</code> for an async template uses
<code>asyncio.run</code>.
:pr:<code>1952</code></li>
<li>Avoid unclosed <code>auto_aiter</code> warnings.
:pr:<code>1960</code></li>
<li>Return an <code>aclose</code>-able <code>AsyncGenerator</code> from
<code>Template.generate_async</code>. :pr:<code>1960</code></li>
<li>Avoid leaving <code>root_render_func()</code> unclosed in
<code>Template.generate_async</code>. :pr:<code>1960</code></li>
<li>Avoid leaving async generators unclosed in blocks, includes and
extends.
:pr:<code>1960</code></li>
<li>The runtime uses the correct <code>concat</code> function for the
current environment
when calling block references. :issue:<code>1701</code></li>
<li>Make <code>|unique</code> async-aware, allowing it to be used after
another
async-aware filter. :issue:<code>1781</code></li>
<li><code>|int</code> filter handles <code>OverflowError</code> from
scientific notation.
:issue:<code>1921</code></li>
<li>Make compiling deterministic for tuple unpacking in a <code>{% set
... %}</code>
call. :issue:<code>2021</code></li>
<li>Fix dunder protocol (<code>copy</code>/<code>pickle</code>/etc)
interaction with <code>Undefined</code>
objects. :issue:<code>2025</code></li>
<li>Fix <code>copy</code>/<code>pickle</code> support for the internal
<code>missing</code> object.
:issue:<code>2027</code></li>
<li><code>Environment.overlay(enable_async)</code> is applied correctly.
:pr:<code>2061</code></li>
<li>The error message from <code>FileSystemLoader</code> includes the
paths that were
searched. :issue:<code>1661</code></li>
<li><code>PackageLoader</code> shows a clearer error message when the
package does not
contain the templates directory. :issue:<code>1705</code></li>
<li>Improve annotations for methods returning copies.
:pr:<code>1880</code></li>
<li><code>urlize</code> does not add <code>mailto:</code> to values like
<code>@a@b</code>. :pr:<code>1870</code></li>
<li>Tests decorated with <code>@pass_context`` can be used with the
``|select`` filter. :issue:</code>1624`</li>
<li>Using <code>set</code> for multiple assignment (<code>a, b = 1,
2</code>) does not fail when the
target is a namespace attribute. :issue:<code>1413</code></li>
<li>Using <code>set</code> in all branches of <code>{% if %}{% elif %}{%
else %}</code> blocks
does not cause the variable to be considered initially undefined.
:issue:<code>1253</code></li>
</ul>
</blockquote>
</details>
<details>
<summary>Commits</summary>
<ul>
<li><a
href="https://github.com/pallets/jinja/commit/877f6e51be8e1765b06d911cfaa9033775f051d1"><code>877f6e5</code></a>
release version 3.1.5</li>
<li><a
href="https://github.com/pallets/jinja/commit/8d588592653b052f957b720e1fc93196e06f207f"><code>8d58859</code></a>
remove test pypi</li>
<li><a
href="https://github.com/pallets/jinja/commit/eda8fe86fd716dfce24910294e9f1fc81fbc740c"><code>eda8fe8</code></a>
update dev dependencies</li>
<li><a
href="https://github.com/pallets/jinja/commit/c8fdce1e0333f1122b244b03a48535fdd7b03d91"><code>c8fdce1</code></a>
Fix bug involving calling set on a template parameter within all
branches of ...</li>
<li><a
href="https://github.com/pallets/jinja/commit/66587ce989e5a478e0bb165371fa2b9d42b7040f"><code>66587ce</code></a>
Fix bug where set would sometimes fail within if</li>
<li><a
href="https://github.com/pallets/jinja/commit/fbc3a696c729d177340cc089531de7e2e5b6f065"><code>fbc3a69</code></a>
Add support for namespaces in tuple parsing (<a
href="https://redirect.github.com/pallets/jinja/issues/1664">#1664</a>)</li>
<li><a
href="https://github.com/pallets/jinja/commit/b8f4831d41e6a7cb5c40d42f074ffd92d2daccfc"><code>b8f4831</code></a>
more comments about nsref assignment</li>
<li><a
href="https://github.com/pallets/jinja/commit/ee832194cd9f55f75e5a51359b709d535efe957f"><code>ee83219</code></a>
Add support for namespaces in tuple assignment</li>
<li><a
href="https://github.com/pallets/jinja/commit/1d55cddbb28e433779511f28f13a2d8c4ec45826"><code>1d55cdd</code></a>
Triple quotes in docs (<a
href="https://redirect.github.com/pallets/jinja/issues/2064">#2064</a>)</li>
<li><a
href="https://github.com/pallets/jinja/commit/8a8eafc6b992ba177f1d3dd483f8465f18a11116"><code>8a8eafc</code></a>
edit block assignment section</li>
<li>Additional commits viewable in <a
href="https://github.com/pallets/jinja/compare/3.1.4...3.1.5">compare
view</a></li>
</ul>
</details>
<br />
[](https://docs.github.com/en/github/managing-security-vulnerabilities/about-dependabot-security-updates#about-compatibility-scores)
Dependabot will resolve any conflicts with this PR as long as you don't
alter it yourself. You can also trigger a rebase manually by commenting
`@dependabot rebase`.
[//]: # (dependabot-automerge-start)
[//]: # (dependabot-automerge-end)
---
<details>
<summary>Dependabot commands and options</summary>
<br />
You can trigger Dependabot actions by commenting on this PR:
- `@dependabot rebase` will rebase this PR
- `@dependabot recreate` will recreate this PR, overwriting any edits
that have been made to it
- `@dependabot merge` will merge this PR after your CI passes on it
- `@dependabot squash and merge` will squash and merge this PR after
your CI passes on it
- `@dependabot cancel merge` will cancel a previously requested merge
and block automerging
- `@dependabot reopen` will reopen this PR if it is closed
- `@dependabot close` will close this PR and stop Dependabot recreating
it. You can achieve the same result by closing it manually
- `@dependabot show <dependency name> ignore conditions` will show all
of the ignore conditions of the specified dependency
- `@dependabot ignore this major version` will close this PR and stop
Dependabot creating any more for this major version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this minor version` will close this PR and stop
Dependabot creating any more for this minor version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this dependency` will close this PR and stop
Dependabot creating any more for this dependency (unless you reopen the
PR or upgrade to it yourself)
You can disable automated security fix PRs for this repo from the
[Security Alerts
page](https://github.com/langchain-ai/langgraph/network/alerts).
</details>
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
When looking at [docs](https://langchain-ai.github.io/langgraph/) this
sentence is confusing, not clear there's two separate links or why one
of them would lead to repo
Bumps [jinja2](https://github.com/pallets/jinja) from 3.1.4 to 3.1.5.
<details>
<summary>Release notes</summary>
<p><em>Sourced from <a
href="https://github.com/pallets/jinja/releases">jinja2's
releases</a>.</em></p>
<blockquote>
<h2>3.1.5</h2>
<p>This is the Jinja 3.1.5 security fix release, which fixes security
issues and bugs but does not otherwise change behavior and should not
result in breaking changes compared to the latest feature release.</p>
<p>PyPI: <a
href="https://pypi.org/project/Jinja2/3.1.5/">https://pypi.org/project/Jinja2/3.1.5/</a>
Changes: <a
href="https://jinja.palletsprojects.com/changes/#version-3-1-5">https://jinja.palletsprojects.com/changes/#version-3-1-5</a>
Milestone: <a
href="https://github.com/pallets/jinja/milestone/16?closed=1">https://github.com/pallets/jinja/milestone/16?closed=1</a></p>
<ul>
<li>The sandboxed environment handles indirect calls to
<code>str.format</code>, such as by passing a stored reference to a
filter that calls its argument. <a
href="https://github.com/pallets/jinja/security/advisories/GHSA-q2x7-8rv6-6q7h">GHSA-q2x7-8rv6-6q7h</a></li>
<li>Escape template name before formatting it into error messages, to
avoid issues with names that contain f-string syntax. <a
href="https://redirect.github.com/pallets/jinja/issues/1792">#1792</a>,
<a
href="https://github.com/pallets/jinja/security/advisories/GHSA-gmj6-6f8f-6699">GHSA-gmj6-6f8f-6699</a></li>
<li>Sandbox does not allow <code>clear</code> and <code>pop</code> on
known mutable sequence types. <a
href="https://redirect.github.com/pallets/jinja/issues/2032">#2032</a></li>
<li>Calling sync <code>render</code> for an async template uses
<code>asyncio.run</code>. <a
href="https://redirect.github.com/pallets/jinja/issues/1952">#1952</a></li>
<li>Avoid unclosed <code>auto_aiter</code> warnings. <a
href="https://redirect.github.com/pallets/jinja/issues/1960">#1960</a></li>
<li>Return an <code>aclose</code>-able <code>AsyncGenerator</code> from
<code>Template.generate_async</code>. <a
href="https://redirect.github.com/pallets/jinja/issues/1960">#1960</a></li>
<li>Avoid leaving <code>root_render_func()</code> unclosed in
<code>Template.generate_async</code>. <a
href="https://redirect.github.com/pallets/jinja/issues/1960">#1960</a></li>
<li>Avoid leaving async generators unclosed in blocks, includes and
extends. <a
href="https://redirect.github.com/pallets/jinja/issues/1960">#1960</a></li>
<li>The runtime uses the correct <code>concat</code> function for the
current environment when calling block references. <a
href="https://redirect.github.com/pallets/jinja/issues/1701">#1701</a></li>
<li>Make <code>|unique</code> async-aware, allowing it to be used after
another async-aware filter. <a
href="https://redirect.github.com/pallets/jinja/issues/1781">#1781</a></li>
<li><code>|int</code> filter handles <code>OverflowError</code> from
scientific notation. <a
href="https://redirect.github.com/pallets/jinja/issues/1921">#1921</a></li>
<li>Make compiling deterministic for tuple unpacking in a <code>{% set
... %}</code> call. <a
href="https://redirect.github.com/pallets/jinja/issues/2021">#2021</a></li>
<li>Fix dunder protocol (<code>copy</code>/<code>pickle</code>/etc)
interaction with <code>Undefined</code> objects. <a
href="https://redirect.github.com/pallets/jinja/issues/2025">#2025</a></li>
<li>Fix <code>copy</code>/<code>pickle</code> support for the internal
<code>missing</code> object. <a
href="https://redirect.github.com/pallets/jinja/issues/2027">#2027</a></li>
<li><code>Environment.overlay(enable_async)</code> is applied correctly.
<a
href="https://redirect.github.com/pallets/jinja/issues/2061">#2061</a></li>
<li>The error message from <code>FileSystemLoader</code> includes the
paths that were searched. <a
href="https://redirect.github.com/pallets/jinja/issues/1661">#1661</a></li>
<li><code>PackageLoader</code> shows a clearer error message when the
package does not contain the templates directory. <a
href="https://redirect.github.com/pallets/jinja/issues/1705">#1705</a></li>
<li>Improve annotations for methods returning copies. <a
href="https://redirect.github.com/pallets/jinja/issues/1880">#1880</a></li>
<li><code>urlize</code> does not add <code>mailto:</code> to values like
<code>@a@b</code>. <a
href="https://redirect.github.com/pallets/jinja/issues/1870">#1870</a></li>
<li>Tests decorated with <code>@pass_context</code> can be used with the
<code>|select</code> filter. <a
href="https://redirect.github.com/pallets/jinja/issues/1624">#1624</a></li>
<li>Using <code>set</code> for multiple assignment (<code>a, b = 1,
2</code>) does not fail when the target is a namespace attribute. <a
href="https://redirect.github.com/pallets/jinja/issues/1413">#1413</a></li>
<li>Using <code>set</code> in all branches of <code>{% if %}{% elif %}{%
else %}</code> blocks does not cause the variable to be considered
initially undefined. <a
href="https://redirect.github.com/pallets/jinja/issues/1253">#1253</a></li>
</ul>
</blockquote>
</details>
<details>
<summary>Changelog</summary>
<p><em>Sourced from <a
href="https://github.com/pallets/jinja/blob/main/CHANGES.rst">jinja2's
changelog</a>.</em></p>
<blockquote>
<h2>Version 3.1.5</h2>
<p>Released 2024-12-21</p>
<ul>
<li>The sandboxed environment handles indirect calls to
<code>str.format</code>, such as
by passing a stored reference to a filter that calls its argument.
:ghsa:<code>q2x7-8rv6-6q7h</code></li>
<li>Escape template name before formatting it into error messages, to
avoid
issues with names that contain f-string syntax.
:issue:<code>1792</code>, :ghsa:<code>gmj6-6f8f-6699</code></li>
<li>Sandbox does not allow <code>clear</code> and <code>pop</code> on
known mutable sequence
types. :issue:<code>2032</code></li>
<li>Calling sync <code>render</code> for an async template uses
<code>asyncio.run</code>.
:pr:<code>1952</code></li>
<li>Avoid unclosed <code>auto_aiter</code> warnings.
:pr:<code>1960</code></li>
<li>Return an <code>aclose</code>-able <code>AsyncGenerator</code> from
<code>Template.generate_async</code>. :pr:<code>1960</code></li>
<li>Avoid leaving <code>root_render_func()</code> unclosed in
<code>Template.generate_async</code>. :pr:<code>1960</code></li>
<li>Avoid leaving async generators unclosed in blocks, includes and
extends.
:pr:<code>1960</code></li>
<li>The runtime uses the correct <code>concat</code> function for the
current environment
when calling block references. :issue:<code>1701</code></li>
<li>Make <code>|unique</code> async-aware, allowing it to be used after
another
async-aware filter. :issue:<code>1781</code></li>
<li><code>|int</code> filter handles <code>OverflowError</code> from
scientific notation.
:issue:<code>1921</code></li>
<li>Make compiling deterministic for tuple unpacking in a <code>{% set
... %}</code>
call. :issue:<code>2021</code></li>
<li>Fix dunder protocol (<code>copy</code>/<code>pickle</code>/etc)
interaction with <code>Undefined</code>
objects. :issue:<code>2025</code></li>
<li>Fix <code>copy</code>/<code>pickle</code> support for the internal
<code>missing</code> object.
:issue:<code>2027</code></li>
<li><code>Environment.overlay(enable_async)</code> is applied correctly.
:pr:<code>2061</code></li>
<li>The error message from <code>FileSystemLoader</code> includes the
paths that were
searched. :issue:<code>1661</code></li>
<li><code>PackageLoader</code> shows a clearer error message when the
package does not
contain the templates directory. :issue:<code>1705</code></li>
<li>Improve annotations for methods returning copies.
:pr:<code>1880</code></li>
<li><code>urlize</code> does not add <code>mailto:</code> to values like
<code>@a@b</code>. :pr:<code>1870</code></li>
<li>Tests decorated with <code>@pass_context`` can be used with the
``|select`` filter. :issue:</code>1624`</li>
<li>Using <code>set</code> for multiple assignment (<code>a, b = 1,
2</code>) does not fail when the
target is a namespace attribute. :issue:<code>1413</code></li>
<li>Using <code>set</code> in all branches of <code>{% if %}{% elif %}{%
else %}</code> blocks
does not cause the variable to be considered initially undefined.
:issue:<code>1253</code></li>
</ul>
</blockquote>
</details>
<details>
<summary>Commits</summary>
<ul>
<li><a
href="https://github.com/pallets/jinja/commit/877f6e51be8e1765b06d911cfaa9033775f051d1"><code>877f6e5</code></a>
release version 3.1.5</li>
<li><a
href="https://github.com/pallets/jinja/commit/8d588592653b052f957b720e1fc93196e06f207f"><code>8d58859</code></a>
remove test pypi</li>
<li><a
href="https://github.com/pallets/jinja/commit/eda8fe86fd716dfce24910294e9f1fc81fbc740c"><code>eda8fe8</code></a>
update dev dependencies</li>
<li><a
href="https://github.com/pallets/jinja/commit/c8fdce1e0333f1122b244b03a48535fdd7b03d91"><code>c8fdce1</code></a>
Fix bug involving calling set on a template parameter within all
branches of ...</li>
<li><a
href="https://github.com/pallets/jinja/commit/66587ce989e5a478e0bb165371fa2b9d42b7040f"><code>66587ce</code></a>
Fix bug where set would sometimes fail within if</li>
<li><a
href="https://github.com/pallets/jinja/commit/fbc3a696c729d177340cc089531de7e2e5b6f065"><code>fbc3a69</code></a>
Add support for namespaces in tuple parsing (<a
href="https://redirect.github.com/pallets/jinja/issues/1664">#1664</a>)</li>
<li><a
href="https://github.com/pallets/jinja/commit/b8f4831d41e6a7cb5c40d42f074ffd92d2daccfc"><code>b8f4831</code></a>
more comments about nsref assignment</li>
<li><a
href="https://github.com/pallets/jinja/commit/ee832194cd9f55f75e5a51359b709d535efe957f"><code>ee83219</code></a>
Add support for namespaces in tuple assignment</li>
<li><a
href="https://github.com/pallets/jinja/commit/1d55cddbb28e433779511f28f13a2d8c4ec45826"><code>1d55cdd</code></a>
Triple quotes in docs (<a
href="https://redirect.github.com/pallets/jinja/issues/2064">#2064</a>)</li>
<li><a
href="https://github.com/pallets/jinja/commit/8a8eafc6b992ba177f1d3dd483f8465f18a11116"><code>8a8eafc</code></a>
edit block assignment section</li>
<li>Additional commits viewable in <a
href="https://github.com/pallets/jinja/compare/3.1.4...3.1.5">compare
view</a></li>
</ul>
</details>
<br />
[](https://docs.github.com/en/github/managing-security-vulnerabilities/about-dependabot-security-updates#about-compatibility-scores)
Dependabot will resolve any conflicts with this PR as long as you don't
alter it yourself. You can also trigger a rebase manually by commenting
`@dependabot rebase`.
[//]: # (dependabot-automerge-start)
[//]: # (dependabot-automerge-end)
---
<details>
<summary>Dependabot commands and options</summary>
<br />
You can trigger Dependabot actions by commenting on this PR:
- `@dependabot rebase` will rebase this PR
- `@dependabot recreate` will recreate this PR, overwriting any edits
that have been made to it
- `@dependabot merge` will merge this PR after your CI passes on it
- `@dependabot squash and merge` will squash and merge this PR after
your CI passes on it
- `@dependabot cancel merge` will cancel a previously requested merge
and block automerging
- `@dependabot reopen` will reopen this PR if it is closed
- `@dependabot close` will close this PR and stop Dependabot recreating
it. You can achieve the same result by closing it manually
- `@dependabot show <dependency name> ignore conditions` will show all
of the ignore conditions of the specified dependency
- `@dependabot ignore this major version` will close this PR and stop
Dependabot creating any more for this major version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this minor version` will close this PR and stop
Dependabot creating any more for this minor version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this dependency` will close this PR and stop
Dependabot creating any more for this dependency (unless you reopen the
PR or upgrade to it yourself)
You can disable automated security fix PRs for this repo from the
[Security Alerts
page](https://github.com/langchain-ai/langgraph/network/alerts).
</details>
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Within `libs/langgraph`, change all `TypedDict` imports to come from
`typing_extensions` rather than `typing`, as `pydantic` doesn't like the
latter.
Additionally, add a ruff rule to ban these imports too (so this doesn't
regress).
Solves #2909.
This PR adds a "shallow" version of `PostgresSaver` checkpointer that
ONLY stores the most recent checkpoint and does NOT retain any history.
It is meant to be a light-weight drop-in replacement for the
PostgresSaver that supports most of the LangGraph persistence
functionality with the exception of time travel.
Made some of the explanations more clear by rephrasing certain parts of
the sentence.
Fixed minor grammar mistakes also.
---------
Co-authored-by: Vadym Barda <vadim.barda@gmail.com>
Bumps [tornado](https://github.com/tornadoweb/tornado) from 6.4.1 to
6.4.2.
<details>
<summary>Changelog</summary>
<p><em>Sourced from <a
href="https://github.com/tornadoweb/tornado/blob/v6.4.2/docs/releases.rst">tornado's
changelog</a>.</em></p>
<blockquote>
<h1>Release notes</h1>
<p>.. toctree::
:maxdepth: 2</p>
<p>releases/v6.4.2
releases/v6.4.1
releases/v6.4.0
releases/v6.3.3
releases/v6.3.2
releases/v6.3.1
releases/v6.3.0
releases/v6.2.0
releases/v6.1.0
releases/v6.0.4
releases/v6.0.3
releases/v6.0.2
releases/v6.0.1
releases/v6.0.0
releases/v5.1.1
releases/v5.1.0
releases/v5.0.2
releases/v5.0.1
releases/v5.0.0
releases/v4.5.3
releases/v4.5.2
releases/v4.5.1
releases/v4.5.0
releases/v4.4.3
releases/v4.4.2
releases/v4.4.1
releases/v4.4.0
releases/v4.3.0
releases/v4.2.1
releases/v4.2.0
releases/v4.1.0
releases/v4.0.2
releases/v4.0.1
releases/v4.0.0
releases/v3.2.2
releases/v3.2.1
releases/v3.2.0
releases/v3.1.1
releases/v3.1.0
releases/v3.0.2
releases/v3.0.1
releases/v3.0.0
releases/v2.4.1
releases/v2.4.0</p>
<!-- raw HTML omitted -->
</blockquote>
<p>... (truncated)</p>
</details>
<details>
<summary>Commits</summary>
<ul>
<li><a
href="https://github.com/tornadoweb/tornado/commit/a5ecfab15e52202a46d34638aad93cddca86d87b"><code>a5ecfab</code></a>
Bump version to 6.4.2</li>
<li><a
href="https://github.com/tornadoweb/tornado/commit/bc7df6bafdec61155e7bf385081feb205463857d"><code>bc7df6b</code></a>
Fix tests with Twisted 24.7.0</li>
<li><a
href="https://github.com/tornadoweb/tornado/commit/d5ba4a1695fbf7c6a3e54313262639b198291533"><code>d5ba4a1</code></a>
httputil: Fix quadratic performance of cookie parsing</li>
<li>See full diff in <a
href="https://github.com/tornadoweb/tornado/compare/v6.4.1...v6.4.2">compare
view</a></li>
</ul>
</details>
<br />
[](https://docs.github.com/en/github/managing-security-vulnerabilities/about-dependabot-security-updates#about-compatibility-scores)
Dependabot will resolve any conflicts with this PR as long as you don't
alter it yourself. You can also trigger a rebase manually by commenting
`@dependabot rebase`.
[//]: # (dependabot-automerge-start)
[//]: # (dependabot-automerge-end)
---
<details>
<summary>Dependabot commands and options</summary>
<br />
You can trigger Dependabot actions by commenting on this PR:
- `@dependabot rebase` will rebase this PR
- `@dependabot recreate` will recreate this PR, overwriting any edits
that have been made to it
- `@dependabot merge` will merge this PR after your CI passes on it
- `@dependabot squash and merge` will squash and merge this PR after
your CI passes on it
- `@dependabot cancel merge` will cancel a previously requested merge
and block automerging
- `@dependabot reopen` will reopen this PR if it is closed
- `@dependabot close` will close this PR and stop Dependabot recreating
it. You can achieve the same result by closing it manually
- `@dependabot show <dependency name> ignore conditions` will show all
of the ignore conditions of the specified dependency
- `@dependabot ignore this major version` will close this PR and stop
Dependabot creating any more for this major version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this minor version` will close this PR and stop
Dependabot creating any more for this minor version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this dependency` will close this PR and stop
Dependabot creating any more for this dependency (unless you reopen the
PR or upgrade to it yourself)
You can disable automated security fix PRs for this repo from the
[Security Alerts
page](https://github.com/langchain-ai/langgraph/network/alerts).
</details>
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Vadym Barda <vadym@langchain.dev>
Config parameter in `dev` was typed as pathlib.Path, but it is actually
a string. We need to manually create a Path from the string when parsing
the config.
FIxes#2647
Add `format` flag to `add_messages` which allows you to specify if the
contents of messages in state should be formatted in a particular way.
PR only adds support for OpenAI style contents. Helpful if you're using
different models at different nodes and want a unified messages format
to interact with when you manually update messages.
Bumps [tornado](https://github.com/tornadoweb/tornado) from 6.4.1 to
6.4.2.
<details>
<summary>Changelog</summary>
<p><em>Sourced from <a
href="https://github.com/tornadoweb/tornado/blob/v6.4.2/docs/releases.rst">tornado's
changelog</a>.</em></p>
<blockquote>
<h1>Release notes</h1>
<p>.. toctree::
:maxdepth: 2</p>
<p>releases/v6.4.2
releases/v6.4.1
releases/v6.4.0
releases/v6.3.3
releases/v6.3.2
releases/v6.3.1
releases/v6.3.0
releases/v6.2.0
releases/v6.1.0
releases/v6.0.4
releases/v6.0.3
releases/v6.0.2
releases/v6.0.1
releases/v6.0.0
releases/v5.1.1
releases/v5.1.0
releases/v5.0.2
releases/v5.0.1
releases/v5.0.0
releases/v4.5.3
releases/v4.5.2
releases/v4.5.1
releases/v4.5.0
releases/v4.4.3
releases/v4.4.2
releases/v4.4.1
releases/v4.4.0
releases/v4.3.0
releases/v4.2.1
releases/v4.2.0
releases/v4.1.0
releases/v4.0.2
releases/v4.0.1
releases/v4.0.0
releases/v3.2.2
releases/v3.2.1
releases/v3.2.0
releases/v3.1.1
releases/v3.1.0
releases/v3.0.2
releases/v3.0.1
releases/v3.0.0
releases/v2.4.1
releases/v2.4.0</p>
<!-- raw HTML omitted -->
</blockquote>
<p>... (truncated)</p>
</details>
<details>
<summary>Commits</summary>
<ul>
<li><a
href="https://github.com/tornadoweb/tornado/commit/a5ecfab15e52202a46d34638aad93cddca86d87b"><code>a5ecfab</code></a>
Bump version to 6.4.2</li>
<li><a
href="https://github.com/tornadoweb/tornado/commit/bc7df6bafdec61155e7bf385081feb205463857d"><code>bc7df6b</code></a>
Fix tests with Twisted 24.7.0</li>
<li><a
href="https://github.com/tornadoweb/tornado/commit/d5ba4a1695fbf7c6a3e54313262639b198291533"><code>d5ba4a1</code></a>
httputil: Fix quadratic performance of cookie parsing</li>
<li>See full diff in <a
href="https://github.com/tornadoweb/tornado/compare/v6.4.1...v6.4.2">compare
view</a></li>
</ul>
</details>
<br />
[](https://docs.github.com/en/github/managing-security-vulnerabilities/about-dependabot-security-updates#about-compatibility-scores)
Dependabot will resolve any conflicts with this PR as long as you don't
alter it yourself. You can also trigger a rebase manually by commenting
`@dependabot rebase`.
[//]: # (dependabot-automerge-start)
[//]: # (dependabot-automerge-end)
---
<details>
<summary>Dependabot commands and options</summary>
<br />
You can trigger Dependabot actions by commenting on this PR:
- `@dependabot rebase` will rebase this PR
- `@dependabot recreate` will recreate this PR, overwriting any edits
that have been made to it
- `@dependabot merge` will merge this PR after your CI passes on it
- `@dependabot squash and merge` will squash and merge this PR after
your CI passes on it
- `@dependabot cancel merge` will cancel a previously requested merge
and block automerging
- `@dependabot reopen` will reopen this PR if it is closed
- `@dependabot close` will close this PR and stop Dependabot recreating
it. You can achieve the same result by closing it manually
- `@dependabot show <dependency name> ignore conditions` will show all
of the ignore conditions of the specified dependency
- `@dependabot ignore this major version` will close this PR and stop
Dependabot creating any more for this major version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this minor version` will close this PR and stop
Dependabot creating any more for this minor version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this dependency` will close this PR and stop
Dependabot creating any more for this dependency (unless you reopen the
PR or upgrade to it yourself)
You can disable automated security fix PRs for this repo from the
[Security Alerts
page](https://github.com/langchain-ai/langgraph/network/alerts).
</details>
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Vadym Barda <vadym@langchain.dev>
Add `format` flag to `add_messages` which allows you to specify if the
contents of messages in state should be formatted in a particular way.
PR only adds support for OpenAI style contents. Helpful if you're using
different models at different nodes and want a unified messages format
to interact with when you manually update messages.
Bumps [tornado](https://github.com/tornadoweb/tornado) from 6.4.1 to
6.4.2.
<details>
<summary>Changelog</summary>
<p><em>Sourced from <a
href="https://github.com/tornadoweb/tornado/blob/v6.4.2/docs/releases.rst">tornado's
changelog</a>.</em></p>
<blockquote>
<h1>Release notes</h1>
<p>.. toctree::
:maxdepth: 2</p>
<p>releases/v6.4.2
releases/v6.4.1
releases/v6.4.0
releases/v6.3.3
releases/v6.3.2
releases/v6.3.1
releases/v6.3.0
releases/v6.2.0
releases/v6.1.0
releases/v6.0.4
releases/v6.0.3
releases/v6.0.2
releases/v6.0.1
releases/v6.0.0
releases/v5.1.1
releases/v5.1.0
releases/v5.0.2
releases/v5.0.1
releases/v5.0.0
releases/v4.5.3
releases/v4.5.2
releases/v4.5.1
releases/v4.5.0
releases/v4.4.3
releases/v4.4.2
releases/v4.4.1
releases/v4.4.0
releases/v4.3.0
releases/v4.2.1
releases/v4.2.0
releases/v4.1.0
releases/v4.0.2
releases/v4.0.1
releases/v4.0.0
releases/v3.2.2
releases/v3.2.1
releases/v3.2.0
releases/v3.1.1
releases/v3.1.0
releases/v3.0.2
releases/v3.0.1
releases/v3.0.0
releases/v2.4.1
releases/v2.4.0</p>
<!-- raw HTML omitted -->
</blockquote>
<p>... (truncated)</p>
</details>
<details>
<summary>Commits</summary>
<ul>
<li><a
href="https://github.com/tornadoweb/tornado/commit/a5ecfab15e52202a46d34638aad93cddca86d87b"><code>a5ecfab</code></a>
Bump version to 6.4.2</li>
<li><a
href="https://github.com/tornadoweb/tornado/commit/bc7df6bafdec61155e7bf385081feb205463857d"><code>bc7df6b</code></a>
Fix tests with Twisted 24.7.0</li>
<li><a
href="https://github.com/tornadoweb/tornado/commit/d5ba4a1695fbf7c6a3e54313262639b198291533"><code>d5ba4a1</code></a>
httputil: Fix quadratic performance of cookie parsing</li>
<li>See full diff in <a
href="https://github.com/tornadoweb/tornado/compare/v6.4.1...v6.4.2">compare
view</a></li>
</ul>
</details>
<br />
[](https://docs.github.com/en/github/managing-security-vulnerabilities/about-dependabot-security-updates#about-compatibility-scores)
Dependabot will resolve any conflicts with this PR as long as you don't
alter it yourself. You can also trigger a rebase manually by commenting
`@dependabot rebase`.
[//]: # (dependabot-automerge-start)
[//]: # (dependabot-automerge-end)
---
<details>
<summary>Dependabot commands and options</summary>
<br />
You can trigger Dependabot actions by commenting on this PR:
- `@dependabot rebase` will rebase this PR
- `@dependabot recreate` will recreate this PR, overwriting any edits
that have been made to it
- `@dependabot merge` will merge this PR after your CI passes on it
- `@dependabot squash and merge` will squash and merge this PR after
your CI passes on it
- `@dependabot cancel merge` will cancel a previously requested merge
and block automerging
- `@dependabot reopen` will reopen this PR if it is closed
- `@dependabot close` will close this PR and stop Dependabot recreating
it. You can achieve the same result by closing it manually
- `@dependabot show <dependency name> ignore conditions` will show all
of the ignore conditions of the specified dependency
- `@dependabot ignore this major version` will close this PR and stop
Dependabot creating any more for this major version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this minor version` will close this PR and stop
Dependabot creating any more for this minor version (unless you reopen
the PR or upgrade to it yourself)
- `@dependabot ignore this dependency` will close this PR and stop
Dependabot creating any more for this dependency (unless you reopen the
PR or upgrade to it yourself)
You can disable automated security fix PRs for this repo from the
[Security Alerts
page](https://github.com/langchain-ai/langgraph/network/alerts).
</details>
Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
Co-authored-by: Vadym Barda <vadym@langchain.dev>
Config parameter in `dev` was typed as pathlib.Path, but it is actually
a string. We need to manually create a Path from the string when parsing
the config.
Hi,
While reading the [update state from
tools](https://langchain-ai.github.io/langgraph/how-tos/update-state-from-tools/)
tutorial. I noticed that this code snippet contains a syntax error:
```python
def call_tools(state):
...
commands = [tools_by_name[call["name"].invoke(call, config={"coerce_tool_content": False}) for tool_call in tool_calls]
return commands
```
There is a missing closing bracket `]` in the list comprehension.
Additionally, the variable `call` inside the list comprehension is
undefined, it should be `tool_call`.
Here is a corrected version of the code:
```python
def call_tools(state):
...
commands = [tools_by_name[tool_call["name"]].invoke(tool_call, config={"coerce_tool_content": False}) for tool_call in tool_calls]
return commands
```
---------
Co-authored-by: Vadym Barda <vadim.barda@gmail.com>
This PR updates the [How-to
guide](https://langchain-ai.github.io/langgraph/how-tos/persistence_mongodb/)
on using the MongoDB checkpointer.
The guide currently explains how to create a custom MongoDB
checkpointer, but we now have a checkpointer implementation available
via the `langgraph-checkpoint-mongodb` library. This PR updates the
current resource to guide users on how to use this implementation.
---------
Co-authored-by: ajosh0504 <apoorva.joshi@mongodb.com>
Co-authored-by: vbarda <vadym@langchain.dev>
- Document interrupt reference
- Update conceptual guides for HIL
- Split time-travel conceptual guide
- Split breakpoints into separate conceptual guide
- Update relevant how-tos
- Update how-to index page for HIL with more information and recommendations
- New how-to for multi turn conversation
- don't create contextvars.Context/asyncio.Task in RunnableSeq (not needed as each step creates it if necessary)
- don't run in-memory-saver methods in background threads (no point as they hold the gil)
- avoid calling should_interrupt when no interrupts set
- Whereas Send is for fire-and-forget type of calls, new `call` and `acall` functions are for flows where you want to wait for the node to finish before doing something else
- Because we return regular python future objects (concurrent.futures.Future or asyncio.Future) all the python primitives for working with futures work, eg. wait, gather, etc
Replace hardcoded database saver class names with `cls` in
`from_conn_string` factory methods to improve subclassing support
## Changes
* Replaced direct class instantiations with `cls(conn)` in
`from_conn_string` classmethods across all database implementations
* Updated both synchronous and asynchronous variants for DuckDB,
PostgreSQL, and SQLite savers
## Why
This refactor makes the database saver classes more extensible by
following Python's convention of using `cls` in class methods. This
enables proper inheritance patterns where subclasses can reuse the
factory methods without needing to override them. Previously, the
hardcoded class names would always instantiate the parent class, even
when called from a subclass.
## Testing
The change is backward compatible and doesn't alter existing
functionality. All existing tests should continue to pass as this is
purely a structural refactoring that preserves the current behavior
while improving extensibility.
## Notes
This PR addresses follow up on comments from #2518 - AsyncPostgresSaver
didn't need to be fixed but many of the other DB saver classes did.
It seems that actually once i moved the operators & other things out,
the query planner does do reasonable things and do sequential scanning
if filtered N < some size but the index otherwise, even with namespace
filtering.
Small change to install the dependencies with `edit` mode so that users
or freshman can see the effect immediately when they change the template
code. As below,
`pip install -e .`
It's very good to evaluate how agent works and easy to test &
re-develop!
---------
Signed-off-by: Mingqi Hu <mingqi.hu@intel.com>
Co-authored-by: William FH <13333726+hinthornw@users.noreply.github.com>
Adds a few of preliminaries:
1. Makes the returned "score" actually the result of the requested
operation (cosine, inner_product, l2)
2. Sorts asc, etc. so that if you were to add an HNSW index (and not
have any WHERE filters), it would be used
3. Drop the inner WHERE statement if no namespace or other filters are
provided. See (2) for why.
I don't yet add an index to the migrations since I think we need to
agree on the right balance to ensure it's actually used in common query
patterns.
- Initializing the store with an 'embedding config' -> this contains the
'dims' (used to create the table) and the encoder object (rn langchain
embeddings object, though that is ......)
- Call setup() -> creates the vector table.
Each document has 1 or more vectors associated with it for each json
path in the embedding config.
Would welcome critique and requests!
Leaving the params as the defaults for pgvector but open to feedback if
you think it's important to be able to more transparently configure that
in setup()
```python
from typing import TypedDict, List, Dict, Any, Optional
from langchain_openai import OpenAIEmbeddings
from langgraph.graph import StateGraph
from langgraph.store.postgres import PostgresStore
emb_config = {
"dims": 1536, # OpenAI embedding dimensions
"embed": OpenAIEmbeddings(model="text-embedding-3-small"),
"distance_type": "cosine",
}
with PostgresStore.from_conn_string(
"postgres://postgres:postgres@localhost:5441",
embedding=emb_config,
) as store:
store.setup()
# Define the state type for our graph
class State(TypedDict):
query: str
results: Optional[List[Dict[str, Any]]]
def put_stuff(state: State) -> State:
docs = [
("doc1", {"text": "red apple in kitchen"}),
("doc2", {"text": "blue car in garage"}),
("doc3", {"text": "green apple on table"}),
]
for key, value in docs:
store.put(("docs",), key, value)
def search_stuff(state: State) -> State:
"""Search for documents using vector similarity."""
results = store.search(("docs",), query=state["query"])
return {"results": results}
builder = StateGraph(State)
builder.add_node(put_stuff)
builder.add_node(search_stuff)
builder.add_edge("__start__", "put_stuff")
builder.add_edge("put_stuff", "search_stuff")
# Compile
with PostgresStore.from_conn_string(
"postgres://postgres:postgres@localhost:5441",
embedding=emb_config,
) as store:
chain = builder.compile(store=store)
result = chain.invoke({"query": "sour apple"})
# Print results
for doc in result["results"]:
print(doc.key)
print(doc.value)
print(doc.response_metadata)
```
- This makes the command bubble up out of the current graph and be handled by the calling graph (the immediate parent)
- This could be extended to support eg. ROOT graph, or some other level
- This is asynchronous, so we shouldn't use for regular writes to the output stream (ie those from PregelLoop)
- For writes from subgraphs / nodes this is fine to use, as we make no guarantees about when those show up anyway
- This should only be used in very specific circunstances, sqlite or
postgres adapters much more appropriate in most circunstances
---------
Co-authored-by: William Fu-Hinthorn <13333726+hinthornw@users.noreply.github.com>
- This works similarly to the input() function from stdlib
- calling it in a node interrupts execution
- invoking the graph with Command(resume=...) will set ... as the return value of interrupt() so that the node can access the "answer" to the "question"
- This PR also starts the work to control the graph on invoke/stream with Command() input, to be continued in a future PR
- Keep old code path for compatibility with existing checkpoints
- Keep a similar order of application of updates, in some cases there will be no visible change
- Update task path for Sends to contain the path of all the parent tasks (multiple parents when a Send task creates another Send)
- That lineage path is used to ensure order of application of updates respects their logical lineage (ie updates from parents always applied before their child tasks)
- Move Interrupt writes to use negative indexes, which allow replacing/shadowing (when task is re-run it may interrupt again, or succeed)
- Runner will now attempt to schedule new Send tasks as soon as the write is received (ie while the originating node is still running)
- Update kafka scheduler to support new Send behavior
- Previously order was enforced in prepare_next_tasks, but that's not a good fit for future features
- This changes order between PULL and PUSH tasks, updates from PUSH tasks will now be applied after updates from PULL tasks
- updates from inside Send tasks are applied in the order the Sends were created, if when you fan out, and have each task write results to a list with reducer, the final list is in the order you used when triggering
- Return Control(update_state=, trigger=, send=) from your nodes instead
- Annotate nodes with Control[Literal["destination"]] to see your graph connections drawn
Added a js-example to show it builds
Adapted integration tests after removing the test CLI command
---------
Co-authored-by: Nuno Campos <nuno@langchain.dev>
* Remove land hand sidebar on most pages
* Cleans up some headings
* Adds error reference information to index (it was already on the
sidebar for the how-to page) -- should probably be its own tab?
* Adds an index page for the reference (so it's easier to link to a main
reference page), alternatively we can set up a redirect from index to
graph
This change expands error-handling functionality of the `ToolNode` by
introducing more options for `handle_tool_errors`. Default behavior of
the `ToolNode` is unchanged -- all errors are handled and wrapped in a
`ToolMessage` to be sent back to LLM.
With this change, users have flexibility to only handle the exceptions
that they need to pass back to the LLM:
* they can specify exceptions to handle by passing a tuple of exceptions
in `handle_tool_errors`
* specify `handle_tool_errors=True/str/callable`
* when `handle_tool_errors` is a callable, the signature will be
inspected and exceptions from the signature will be handled
---------
Co-authored-by: vbarda <vadym@langchain.dev>
- Share step/stop logic with PregelLoop
- Add RemainingSteps value which contains the number of remaining steps
- Switch create_react_agent to use RemainingSteps, so that it behave correctly for return_direct tools
Updated the following guides:
docs/docs/how-tos/streaming-content.ipynb
docs/docs/how-tos/streaming-events-from-within-tools-without-langchain.ipynb
docs/docs/how-tos/streaming-events-from-within-tools.ipynb
docs/docs/how-tos/streaming-tokens-without-langchain.ipynb
Updates the following how to guides
docs/docs/how-tos/disable-streaming.ipynb
docs/docs/how-tos/input_output_schema.ipynb
docs/docs/how-tos/many-tools.ipynb
docs/docs/how-tos/map-reduce.ipynb
docs/docs/how-tos/node-retries.ipynb
docs/docs/how-tos/pass-config-to-tools.ipynb
docs/docs/how-tos/pass_private_state.ipynb
Updates the following how to guides:
docs/docs/how-tos/persistence.ipynb
docs/docs/how-tos/persistence_mongodb.ipynb
docs/docs/how-tos/persistence_postgres.ipynb
docs/docs/how-tos/persistence_redis.ipynb
docs/docs/how-tos/react-agent-from-scratch.ipynb
docs/docs/how-tos/react-agent-structured-output.ipynb
docs/docs/how-tos/recursion-limit.ipynb
docs/docs/how-tos/return-when-recursion-limit-hits.ipynb
docs/docs/how-tos/run-id-langsmith.ipynb
docs/docs/how-tos/state-model.ipynb
Add links to the following how-to guides:
docs/docs/how-tos/async.ipynb
docs/docs/how-tos/branching.ipynb
docs/docs/how-tos/configuration.ipynb
docs/docs/how-tos/create-react-agent-hitl.ipynb
docs/docs/how-tos/create-react-agent-memory.ipynb
docs/docs/how-tos/create-react-agent-system-prompt.ipynb
docs/docs/how-tos/create-react-agent.ipynb
Identified two missing concepts:
1) RunnableConfig in LangChain
2) Unclear where ReAct should link in langgraph
- running callback handler in background thread could potentially lead to ordering issues
- deduping on `id()` could lead to messages being dropped if they reused memory address of a previous chunk
* Update HIL conceptual docs
* Update images
* Update figures and text per feedback
* Update links for how-tos
* Embed img directly in ntbks
* Add dynamic breakpoints
* Update figure in ntbk
* Fix link to dynamic breakpoints
* wip
* actually catch errors
* fixing tool-calling-errors and persistence-redis
* poetry changes
* poetry update
* run tutorials
* only tutorials (testing)
* print errors
* rewoo fixes
* remove customer support because of user input
* skip notebooks programatically
* add back how-tos
* remove redundant if
* add cassetes for tutorials
* remove no execution since it is generated by CI,
* remove multi-agent/usaco
* try to run in parallel
* skip notebooks fix
* remove magic/non-magic cells
* msgpack instead of yaml
* prepare_notebooks change
* ignore msgpack for spelling
* use compression for cassettes
* update spell check
* reset poetry changes
* upgrade packages for latest
* no update
* poetry changes
- adds the ability for nodes (including in subgraphs) to emit chunks directly to the output stream, emitted chunks can have any type
- when stream_mode=custom isnt requested by the caller emitted chunks are ignored
* docs: Clarify exceptions retried by default_retry_on in retry policy tutorial
- Added a remark in the tutorial explaining that the `default_retry_on` function retries on any exception except for the following:
- ValueError
- TypeError
- ArithmeticError
- ImportError
- LookupError
- NameError
- SyntaxError
- RuntimeError
- ReferenceError
- StopIteration
- StopAsyncIteration
- OSError
* http status codes
---------
Co-authored-by: Itay Etelis <itay.etelis@ibm.com>
Co-authored-by: isaac hershenson <ihershenson@hmc.edu>
- Correctly distinguish between exception classes, lists/tuples of exception classes, and callables.
- Add support for lists in `retry_on`, alongside tuples.
- Prevent exception classes from being incorrectly treated as callables.
- Raise a `TypeError` if `retry_on` is of an unsupported type.
Previously, `retry_on` in `RetryPolicy` accepted only exception classes, tuples of exception classes, or callables.
Update the `retry_on` type annotation to include `List[Type[Exception]]`
Changes:
- Updated `retry_on` in `RetryPolicy` to accept `List[Type[Exception]]`.
* Implement serialization with msgpack library
- encode custom python objects with a msgpack extension type, with constructor path string, and args encoded as nested msgpack doc
* Smaller msgpack extension types
* Update lock files
* lock
* Don't delegate to pydantic json
* Fix kafka serde
- should use our serializer to load, as inputs to subgraphs are serialized using it
* Failing test
* Ensure retried subgraphs resume from current point (if any)
* Lint
* Cleanup Test
---------
Co-authored-by: Nuno Campos <nuno@boringbits.io>
* Performance improvements in checkpointer libs
- Use sha1 instead of md5 for hashing (faster in python 3.x)
- Use orjson instead of json for json dumping (sadly can't use for json loading)
* Update tests
* Update
* Use random number instead of hash for get_version_number
* Avoid saving writes for the last task to complete in each step
- only when possible, exceptions for ERROR, INTERRUPT, SEND
* Make Channel.from_checkpoint a regular function
- context manager no longer needed since Context became a managed value
* Use __slots__ for Channels
* Fix for kafka
* Remove unused fil;e
* Add benchmark-fast command for running locally
* Small improvements to jsonplus serializer
* Don't use PregelNode.mapper when schema is a typed dict
- All it would do is create a new copy of same dict
* Avoid copying checkpoint when fetching at beginning of loop
* Fix needs array
* Update tests
* Performance improvements in core library
- Avoid creating new callback manager when received one as arg
- Avoid looking for config when already received one as arg
- Avoid copies of values in ensure_config/merge_configs
- Implement version of ensure_config that accepts multiple configs (avoids calling merge_configs first)
- Avoid calling merge_configs when we only need to attach extra tags/metadata
* Fix
* Fix
* Try again
* Debug ci job
* Fix
* Try again
* Try again
* Try again
* Some more variations
* Attach annotation to first changed file
* Fix
* Re-enable benchmarks
- Define protocol for sync and async producer and consumer
- Accept consumer/producer as init args in Orchestrator/Executor
- If not passed in, create default consumer/producer as before
- await future returned by send() instead of flush()
- use consumer groups by default
- process tasks in batches by default, configurable
- manually commit offsets when batch is processed
- Orchestrator and Executor classes to run LangGraph in a distributed fashion using Kafka as a message bus for communication
- Orchestrator and Executor run on-demand when a new message is published to the topic they listen to
- Orchestrator is responsible for running the Pregel algorithm (deciding next tasks to run) and sending messages to the executor topic
- Executor is responsible for executing each task (node), and sending messages to the orchestrator topic when done
- Use a simpler version of RunnableSequence without tracing serialization
- Remove accepts_run_manager check in RunnableCallable
- Remove creation of ChannelWrite dynamically every time conditional edge runs
- This more closely remembers the environment they're used in, so it's what we should be testing
- Remove unnecessary pytest-asyncio dependency, use anyio pytest plugin instead
- Convert remaining async tests using only memory checkpointer to use all existing ones
- Sends are stored through put_writes, so we don't need to also store them inside checkpoint object
- On reading checkpoint, reconstruct pending_sends from the stored writes
- Convert Context to a ManagedValue
- Add shim for old Context constructor
- Add `runtime` flag for managed values, which, prior to serialization, replaces the value with a placeholder, and replaces it back with the actual value on resuming from checkpoint
* Fix semantics of put_writes/list
- put_writes(error) should not prevent saving future successful if task is retried successfully
- put_writes(writes) should be a no-op if non-error writes already exist for that task (this prevents tasks executed more than once from modifying writes previously saved / acted on)
- checkpoints should not include channel default values (ie those without a version)
- list() should fetch and return writes for each checkpoint
* Lint
* Rm print
* Fix import
* Lint
- Save errors produced by tasks, under pending_writes
- Re-work logic to cancel other tasks when one fails, ready to change for interrupt exception
- Update serializer to handle exceptions
- Update get_state/get_state_history with new return value property "tasks" which contains a richer description of the next tasks, currently with id, name and error (if already ran and errored)
Updated the documentation to describe the chatbot node function's return value in a more Pythonic and enthusiastic way. The description now emphasizes that the function returns a dictionary with the updated messages list neatly tucked under the messages key.
- while the inner graph makes progress it overwrites the partial progress checkpoints, eventually keeping only one for each outer step
- implement parent_config in MemorySaver
- fix edge cases in PregelLoop
This fixes a small mistake in the manage-conversation-history notebook
where the model bound with tools was not used, instead, the original model
was used when invocations occur.
* small changes
* harrison comments
* Update docs/docs/cloud/deployment/test_locally.md
---------
Co-authored-by: Harrison Chase <hw.chase.17@gmail.com>
* Fix example
CC @vbarda
* Progress on tool calling errors
* Update
* Rename
* Format
* Clean up outputs
* Revert
* Use stream instead of invoke for final example
* Fix bug in add_conditional_edges when no path_map is provided
When an instance of a callable class is passed as the path arg to
add_conditional_edges but no path_map is provided, get_type_hints(path) is
called, which raises a TypeError (since get_type_hints only accepts a module,
class, method, or function).
This patch fixes the error by trying to get type hints from path.__call__ first,
which should work for instances of callable classes.
Tested: Added a test that raises TypeError without the fix in this patch but
passes with the fix.
* More defensive, additional test
---------
Co-authored-by: Nuno Campos <nuno@langchain.dev>
Plus:
1. Improve docstrings of add_node
2. Update crosslinking of sqlite and aiosqlite docstrings
3. Fix a bunch of links so we can turn on strict validation
* Change LangGraph Deploy to LangGraph Cloud in CLI reference.
* Update main README to link to Cloud docs. Update How-to Guide link in Cloud index page.
* Create Environments Variable reference page.
* Add Authentication to Conceptual Guide.
* Update setup how-to to refer back to CompiledGraph variable name.
* Update how-to notebooks for double texting.
* Add warning about setting top-level variable for CompiledGraph.
description:Report a bug in LangChain. To report a security issue, please instead use the security option below. For questions, please use the GitHub Discussions.
description:Report a bug in LangGraph. To report a security issue, please instead use the security option below. For questions, please use the GitHub Discussions.
labels:["02 Bug Report"]
body:
- type:markdown
@@ -7,35 +7,29 @@ body:
value:>
Thank you for taking the time to file a bug report.
Use this to report bugs in LangChain.
If you're not certain that your issue is due to a bug in LangChain, please use [GitHub Discussions](https://github.com/langchain-ai/langchain/discussions)
to ask for help with your issue.
Use this to report BUGS in LangGraph. For usage questions, feature requests and general design questions, please use [GitHub Discussions](https://github.com/langchain-ai/langgraph/discussions).
Relevant links to check before filing a bug report to see if your issue has already been reported, fixed or
description:Please confirm and check all the following options.
description:Before submitting this issue, please confirm that you have completed all the steps below by checking each option. These steps help ensure your issue is well-defined, relevant, and actionable.
options:
- label:I added a very descriptive title to this issue.
- label:This is a bug, not a usage question. For questions, please use GitHub Discussions.
required:true
- label:I searched the [LangGraph](https://langchain-ai.github.io/langgraph/)/LangChain documentation with the integrated search.
- label:I added a clear and detailed title that summarizes the issue.
required:true
- label:I used the GitHub search to find a similar question and didn't find it.
- label:I read what a minimal reproducible example is (https://stackoverflow.com/help/minimal-reproducible-example).
required:true
- label:I am sure that this is a bug in LangGraph/LangChain rather than my code.
required:true
- label:I am sure this is better as an issue [rather than a GitHub discussion](https://github.com/langchain-ai/langgraph/discussions/new/choose), since this is a LangGraph bug and not a design question.
- label:I included a self-contained, minimal example that demonstrates the issue INCLUDING all the relevant imports. The code run AS IS to reproduce the issue.
required:true
- type:textarea
id:reproduction
@@ -45,22 +39,14 @@ body:
label:Example Code
description:|
Please add a self-contained, [minimal, reproducible, example](https://stackoverflow.com/help/minimal-reproducible-example) with your use case.
If a maintainer can copy it, run it, and see it right away, there's a much higher chance that you'll be able to get help.
**Important!**
* Reduce your code to the minimum required to reproduce the issue if possible. This makes it much easier for others to help you.
* Avoid screenshots when possible, as they are hard to read and (more importantly) don't allow others to copy-and-paste your code.
placeholder:|
from langchain_core.runnables import RunnableLambda
from langgraph.graph import StateGraph
def bad_code(inputs) -> int:
raise NotImplementedError('For demo purpose')
chain = RunnableLambda(bad_code)
chain.invoke('Hello!')
chain = StateGraph(list)
chain.invoke('Hello!')
render:python
- type:textarea
id:error
@@ -82,7 +68,7 @@ body:
Write a short description telling what you are doing, what you expect to happen, and what is currently happening.
placeholder:|
* I'm trying to use the `langchain` library to do X.
* I'm trying to use the `langgraph` library to do X.
* I expect to see Y.
* Instead, it does Z.
validations:
@@ -92,25 +78,8 @@ body:
attributes:
label:System Info
description:|
Please share your system info with us.
"pip freeze | grep langchain"
platform (windows / linux / mac)
python version
OR if you're on a recent version of langchain-core you can paste the output of:
python -m langchain_core.sys_info
placeholder:|
"pip freeze | grep langchain"
platform
python version
Alternatively, if you're on a recent version of langchain-core you can paste the output of:
python -m langchain_core.sys_info
These will only surface LangChain packages, don't forget to include any other relevant
packages you're using (if you're not sure what's relevant, you can paste the entire output of `pip freeze`).
LangGraph follows a monorepo organization, with the following structure:
-`libs/langgraph` is the main Python library, published to pypi as `langgraph`. This contains the majority of the code for the framework, as well as the majority of the unit tests.
-`libs/checkpoint` , published to pypi as `langgraph-checkpoint` contains the base classes for the persistence layer of langgraph. The two main abstractions are BaseCheckpointSaver (base class for persistence of workflow runs step-by-step) and BaseStore (base class for "long-term memory" operations, offering a key-value interface combined with semantic search over documents, used for persisting information across distinct workflow runs). This library is a dependency of both the main langgraph library, as well as implementations of these storage interfaces for specific databases. This library also contains reference implementations
-`libs/checkpoint-postgres` published to pypi as langgraph-checkpoint-postgres, contains implementations of checkpoint and store backed by postgres. Majority of the test coverage is in `libs/langgraph` in the form of tests that run over all storage implementations in the repo.
-`langgraph-java` contains a Java implementation of the langgraph framework, which is in the early stages of development.
## Feature Overview
langgraph is an orchestration framework (in the style of airflow or temporal) designed for LLM applications, with a focus on streaming output, cyclical and parallel workflows, and interrupt/resume capabilities. Applications built with langgraph are variously called workflows, graphs, cognitive architectures, agents. Key features:
1.**Graph-based Architecture**: Build directed computation graphs with nodes and edges
2.**State Management**: Type-safe state schema with custom reducers and transformations
3.**Human-in-the-loop**: Support for interrupts, checkpoints, and tool call review
4.**Persistence**: Save and resume execution with in-memory or database storage
5.**Streaming**: Multiple modes (values, updates, custom) for real-time feedback
6.**Multi-agent Patterns**: Support for network, supervisor, and hierarchical architectures
## Python Development
### Build/Test/Lint Commands
(in the respective subdirectory)
- Run all tests: `make test`
- Run single test: `make test TEST=path/to/test_file.py::test_function`
Thank you for being interested in contributing to LangGraph!
## General guidelines
Here are some things to keep in mind for all types of contributions:
- Follow the ["fork and pull request"](https://docs.github.com/en/get-started/exploring-projects-on-github/contributing-to-a-project) workflow.
- Fill out the checked-in pull request template when opening pull requests. Note related issues and tag relevant maintainers.
- Ensure your PR passes formatting, linting, and testing checks before requesting a review.
- If you would like comments or feedback, please open an issue or discussion and tag a maintainer.
- Backwards compatibility is key. Your changes must not be breaking, except in case of critical bug and security fixes.
- Look for duplicate PRs or issues that have already been opened before opening a new one.
- Keep scope as isolated as possible. As a general rule, your changes should not affect more than one package at a time.
### Bugfixes
For bug fixes, please open up an issue before proposing a fix to ensure the proposal properly addresses the underlying problem. In general, bug fixes should all have an accompanying unit test that fails before the fix.
### New features
For new features, please start a new [discussion](https://github.com/langchain-ai/langgraph/discussions), where the maintainers will help with scoping out the necessary changes.
## Contribute Documentation
Documentation is a vital part of LangGraph. We welcome both new documentation for new features and
community improvements to our current documentation. Please read the resources below before getting started:
As LangGraph continues to grow, the surface area of documentation required to cover it continues to grow too.
This page provides guidelines for anyone writing documentation for LangGraph, as well as some of our philosophies around organization and structure.
## Philosophy
LangGraph's documentation follows the [Diataxis framework](https://diataxis.fr).
Under this framework, all documentation falls under one of four categories: [Tutorials](#tutorials),
[How-to guides](#how-to-guides),
[References](#references), and [Explanations (aka conceptual guides)](#conceptual-guide).
### Tutorials
Tutorials are lessons that take the reader through a practical activity. Their purpose is to help the user
gain understanding of concepts and how they interact by showing one way to achieve some goal in a hands-on way.
They should **avoid** giving
multiple permutations of ways to achieve that goal in-depth. Choice is burdensome. Instead, they should guide a new user through a recommended path to accomplishing a concrete goal. While the end result of a tutorial does not necessarily need to
be completely production-ready, it should be useful and practically satisfy the goal that you clearly stated in the tutorial's introduction.
To quote the Diataxis website:
> A tutorial serves the user’s *acquisition* of skills and knowledge - their study. Its purpose is not to help the user get something done, but to help them learn.
In LangGraph, these are often higher level guides that show off end-to-end use cases.
Some examples include:
- [Build a Customer Support Bot](https://langchain-ai.github.io/langgraph/tutorials/customer-support/customer-support/)
- [Build a SQL Agent](https://langchain-ai.github.io/langgraph/tutorials/sql-agent/)
Here are some high-level tips on writing a good tutorial:
- Focus on guiding the user to get something done, but keep in mind the end-goal is more to impart principles than to create a perfect production system.
- Be specific, not abstract and follow one path.
- No need to go deeply into alternative approaches, but it’s ok to reference them, ideally with a link to an appropriate how-to guide.
- Get "a point on the board" as soon as possible - something the user can run that outputs something.
- You can iterate and expand afterwards.
- Try to frequently checkpoint at given steps where the user can run code and see progress.
- Focus on results, not technical explanation.
- Crosslink heavily to appropriate conceptual/reference pages
- The first time you mention a LangGraph concept, use its full name (e.g. "human-in-the-loop"), and link to its conceptual/other documentation page.
- It's also helpful to add a prerequisite callout that links to any pages with necessary background information.
- End with a recap/next steps section summarizing what the tutorial covered and future reading, such as related how-to guides.
- Use phrases like "Next we can run X & Y. We will expect Z.". Then afterwards, use language like "Notice Z" that recalls our expectations and directs the reader's attention to the topic we are trying to teach.
- Do not shy away from repetition.
### How-to guides
A how-to guide, as the name implies, demonstrates how to do something discrete and specific.
It should assume that the user is already familiar with underlying concepts, and is trying to solve an immediate problem, but
should still give some background or list the scenarios where the information contained within can be relevant.
They can and should discuss alternatives if one approach may be better than another in certain cases.
To quote the Diataxis website:
> A how-to guide serves the work of the already-competent user, whom you can assume to know what they want to do, and to be able to follow your instructions correctly.
Some examples include:
- [How to add persistence to your graph](https://langchain-ai.github.io/langgraph/how-tos/persistence/)
- [How to view and update past graph state](https://langchain-ai.github.io/langgraph/how-tos/human_in_the_loop/time-travel/)
Here are some high-level tips on writing a good how-to guide:
- Clearly explain what you are guiding the user through at the start
- Assume higher intent than a tutorial and show what the user needs to do to get that task done
- Assume familiarity of concepts, but explain why suggested actions are helpful
- Crosslink heavily to conceptual/reference pages
- Discuss alternatives and responses to real-world tradeoffs that may arise when solving a problem
- Use lots of example code, ideally within complete code blocks that the reader can copy and run.
- End with a recap/next steps section summarizing what the tutorial covered and future reading, such as other related how-to guides
### Conceptual guides
LangGraph's conceptual guides fall under the **Explanation** quadrant of Diataxis. They should cover LangChain terms and concepts
in a more abstract way than how-to guides or tutorials, and should be geared towards curious users interested in
gaining a deeper understanding of the framework. Try to avoid excessively large code examples. The goal here is to
impart perspective to the user rather than to finish a practical project. These guides should cover **why** things work they way they do.
To quote the Diataxis website:
> The perspective of explanation is higher and wider than that of the other types. It does not take the user’s eye-level view, as in a how-to guide, or a close-up view of the machinery, like reference material. Its scope in each case is a topic - “an area of knowledge”, that somehow has to be bounded in a reasonable, meaningful way.
Some examples include:
- [What does it mean to be agentic?](https://langchain-ai.github.io/langgraph/concepts/high_level/)
Here are some high-level tips on writing a good conceptual guide:
- Explain design decisions. Why does concept X exist and why was it designed this way?
- Use analogies and reference other concepts and alternatives
- Avoid blending in too much reference content
- You can and should reference content covered in other guides, but make sure to link to them
### References
References contain detailed, low-level information that describes exactly what functionality exists and how to use it.
In LangGraph, this is mainly our API reference pages, which are populated from docstrings within code.
References pages are generally not read end-to-end, but are consulted as necessary when a user needs to know
how to use something specific.
To quote the Diataxis website:
> The only purpose of a reference guide is to describe, as succinctly as possible, and in an orderly way. Whereas the content of tutorials and how-to guides are led by needs of the user, reference material is led by the product it describes.
Many of the reference pages in LangChain are automatically generated from code,
but here are some high-level tips on writing a good docstring:
- Be concise
- Discuss special cases and deviations from a user's expectations
- Go into detail on required inputs and outputs
- Light details on when one might use the feature are fine, but in-depth details belong in other sections.
Each category serves a distinct purpose and requires a specific approach to writing and structuring the content.
## General guidelines
Here are some other guidelines you should think about when writing and organizing documentation.
We generally do not merge new tutorials from outside contributors without an actue need.
We welcome updates as well as new integration docs, how-tos, and references.
### Avoid duplication
Multiple pages that cover the same material in depth are difficult to maintain and cause confusion. There should
be only one (very rarely two), canonical pages for a given concept or feature. Instead, you should link to other guides.
### Link to other sections
Because sections of the docs do not exist in a vacuum, it is important to link to other sections as often as possible
to allow a developer to learn more about an unfamiliar topic inline.
This includes linking to the API references as well as conceptual sections!
### Be concise
In general, take a less-is-more approach. If a section with a good explanation of a concept already exists, you should link to it rather than
re-explain it, unless the concept you are documenting presents some new wrinkle.
Be concise, including in code samples.
### General style
- Use active voice and present tense whenever possible
- Use examples and code snippets to illustrate concepts and usage
- Use appropriate header levels (`#`, `##`, `###`, etc.) to organize the content hierarchically
- Use fewer cells with more code to make copy/paste easier
- Use bullet points and numbered lists to break down information into easily digestible chunks
- Use tables (especially for **Reference** sections) and diagrams often to present information visually
- Include the table of contents for longer documentation pages to help readers navigate the content, but hide it for shorter pages
## Setup
LangChain documentation consists of two components:
1. Main Documentation: Hosted at [https://langchain-ai.github.io](https://langchain-ai.github.io/langgraph/),
this comprehensive resource serves as the primary user-facing documentation.
It covers a wide array of topics, including tutorials, use cases, integrations,
and more, offering extensive guidance on building with LangGraph.
The content for this documentation lives in the `/docs` directory of the monorepo.
2. In-code Documentation: This is documentation of the codebase itself, which is also
used to generate the externally facing [API Reference](https://langchain-ai.github.io/langgraph/reference/graphs/).
The content for the API reference is autogenerated by scanning the docstrings in the codebase. For this reason we ask that developers document their code well.
We appreciate all contributions to the documentation, whether it be fixing a typo,
adding a new tutorial or example and whether it be in the main documentation or the API Reference.
### 📜 Main Documentation
The content for the main documentation is located in the `/docs` directory of the monorepo.
The documentation is written using a combination of ipython notebooks (`.ipynb` files)
and markdown (`.md` files). The notebooks are converted to markdown
and then built using [MkDocs](https://www.mkdocs.org/).
Feel free to make contributions to the main documentation! 🥰
After modifying the documentation:
1. Run the linting and formatting commands (see below) to ensure that the documentation is well-formatted and free of errors.
2. Optionally build the documentation locally to verify that the changes look good.
3. Make a pull request with the changes.
### ⚒️ Linting and Building Documentation Locally
After writing up the documentation, you may want to lint and build the documentation
locally to ensure that it looks good and is free of errors.
If you're unable to build it locally that's okay as well, as you will be able to
see a preview of the documentation on the pull request page.
From the **monorepo root**, run the following command to install the dependencies:
```bash
poetry install --with docs --no-root
```
#### Building
The code that builds the documentation is located in the `/docs` directory of the monorepo.
Before building the documentation, it is always a good idea to clean the build directory:
```bash
make clean-docs
```
You can build and preview the documentation as outlined below:
```bash
make serve-docs
```
#### Linting
The documentation is linted from the **monorepo root**. To lint it, run the following from there:
```bash
make spellcheck
```
### ️In-code Documentation
The in-code documentation is autogenerated from docstrings.
For the API reference to be useful, the codebase must be well-documented. This means that all functions, classes, and methods should have a docstring that explains what they do, what the arguments are, and what the return value is. This is a good practice in general, but it is especially important for LangChain because the API reference is the primary resource for developers to understand how to use the codebase.
We generally follow the [Google Python Style Guide](https://google.github.io/styleguide/pyguide.html#38-comments-and-docstrings) for docstrings.
Here is an example of a well-documented function:
```python
defmy_function(arg1:int,arg2:str)->float:
"""This is a short description of the function. (It should be a single sentence.)
This is a longer description of the function. It should explain what
the function does, what the arguments are, and what the return value is.
It should wrap at 88 characters.
Examples:
This is a section for examples of how to use the function.
.. code-block:: python
my_function(1, "hello")
Args:
arg1: This is a description of arg1. We do not need to specify the type since
it is already specified in the function signature.
> Looking for the JS version? See the [JS repo](https://github.com/langchain-ai/langgraphjs) and the [JS docs](https://langchain-ai.github.io/langgraphjs/).
## Overview
[LangGraph](https://langchain-ai.github.io/langgraph/) is a library for building stateful, multi-actor applications with LLMs, used to create agent and multi-agent workflows. Compared to other LLM frameworks, it offers these core benefits: cycles, controllability, and persistence. LangGraph allows you to define flows that involve cycles, essential for most agentic architectures, differentiating it from DAG-based solutions. As a very low-level framework, it provides fine-grained control over both the flow and state of your application, crucial for creating reliable agents. Additionally, LangGraph includes built-in persistence, enabling advanced human-in-the-loop and memory features.
[LangGraph](https://langchain-ai.github.io/langgraph/) is a library for building
stateful, multi-actor applications with LLMs, used to create agent and multi-agent
workflows. Check out an introductory tutorial [here](https://langchain-ai.github.io/langgraph/tutorials/introduction/).
LangGraph is inspired by [Pregel](https://research.google/pubs/pub37252/) and [Apache Beam](https://beam.apache.org/). The public interface draws inspiration from [NetworkX](https://networkx.org/documentation/latest/). LangGraph is built by LangChain Inc, the creators of LangChain, but can be used without LangChain.
### Key Features
### Why use LangGraph?
- **Cycles and Branching**: Implement loops and conditionals in your apps.
- **Persistence**: Automatically save state after each step in the graph. Pause and resume the graph execution at any point to support error recovery, human-in-the-loop workflows, time travel and more.
- **Human-in-the-Loop**: Interrupt graph execution to approve or edit next action planned by the agent.
- **Streaming Support**: Stream outputs as they are produced by each node (including token streaming).
- **Integration with LangChain**: LangGraph integrates seamlessly with [LangChain](https://github.com/langchain-ai/langchain/) and [LangSmith](https://docs.smith.langchain.com/) (but does not require them).
LangGraph powers [production-grade agents](https://www.langchain.com/built-with-langgraph), trusted by Linkedin, Uber, Klarna, GitLab, and many more. LangGraph provides fine-grained control over both the flow and state of your agent applications. It implements a central [persistence layer](https://langchain-ai.github.io/langgraph/concepts/persistence/), enabling features that are common to most agent architectures:
- **Memory**: LangGraph persists arbitrary aspects of your application's state,
supporting memory of conversations and other updates within and across user
interactions;
- **Human-in-the-loop**: Because state is checkpointed, execution can be interrupted
and resumed, allowing for decisions, validation, and corrections at key stages via
human input.
Standardizing these components allows individuals and teams to focus on the behavior
of their agent, instead of its supporting infrastructure.
Through [LangGraph Platform](#langgraph-platform), LangGraph also provides tooling for
the development, deployment, debugging, and monitoring of your applications.
LangGraph integrates seamlessly with
[LangChain](https://python.langchain.com/docs/introduction/) and
[LangSmith](https://docs.smith.langchain.com/) (but does not require them).
To learn more about LangGraph, check out our first LangChain Academy
course, *Introduction to LangGraph*, available for free
[LangGraph Platform](https://langchain-ai.github.io/langgraph/concepts/langgraph_platform) is infrastructure for deploying LangGraph agents. It is a commercial solution for deploying agentic applications to production, built on the open-source LangGraph framework. The LangGraph Platform consists of several components that work together to support the development, deployment, debugging, and monitoring of LangGraph applications: [LangGraph Server](https://langchain-ai.github.io/langgraph/concepts/langgraph_server) (APIs), [LangGraph SDKs](https://langchain-ai.github.io/langgraph/concepts/sdk) (clients for the APIs), [LangGraph CLI](https://langchain-ai.github.io/langgraph/concepts/langgraph_cli) (command line tool for building the server), and [LangGraph Studio](https://langchain-ai.github.io/langgraph/concepts/langgraph_studio) (UI/debugger).
See deployment options [here](https://langchain-ai.github.io/langgraph/concepts/deployment_options/)
(includes a free tier).
Here are some common issues that arise in complex deployments, which LangGraph Platform addresses:
- **Streaming support**: LangGraph Server provides [multiple streaming modes](https://langchain-ai.github.io/langgraph/concepts/streaming) optimized for various application needs
- **Background runs**: Runs agents asynchronously in the background
- **Support for long running agents**: Infrastructure that can handle long running processes
- **[Double texting](https://langchain-ai.github.io/langgraph/concepts/double_texting)**: Handle the case where you get two messages from the user before the agent can respond
- **Handle burstiness**: Task queue for ensuring requests are handled consistently without loss, even under heavy loads
## Installation
@@ -31,45 +67,114 @@ pip install -U langgraph
## Example
One of the central concepts of LangGraph is state. Each graph execution creates a state that is passed between nodes in the graph as they execute, and each node updates this internal state with its return value after it executes. The way that the graph updates its internal state is defined by either the type of graph chosen or a custom function.
Let's take a look at a simple example of an agent that can search the web using [Tavily Search API](https://tavily.com/).
Let's build a tool-calling [ReAct-style](https://langchain-ai.github.io/langgraph/concepts/agentic_concepts/#react-implementation) agent that uses a search tool!
```shell
pip install langchain_openai langchain_community
pip install langchain-anthropic
```
```shell
exportOPENAI_API_KEY=sk-...
exportTAVILY_API_KEY=tvly-...
exportANTHROPIC_API_KEY=sk-...
```
Optionally, we can set up [LangSmith](https://docs.smith.langchain.com/) for best-in-class observability.
```shell
exportLANGCHAIN_TRACING_V2="true"
exportLANGCHAIN_API_KEY=ls__...
exportLANGSMITH_TRACING=true
exportLANGSMITH_API_KEY=lsv2_sk_...
```
```python
fromtypingimportAnnotated,Literal,TypedDict
The simplest way to create a tool-calling agent in LangGraph is to use `create_react_agent`:
{"messages":[{"role":"user","content":"what is the weather in sf"}]},
config={"configurable":{"thread_id":42}}
)
final_state["messages"][-1].content
```
```
"Based on the search results, I can tell you that the current weather in San Francisco is:\n\nTemperature: 60 degrees Fahrenheit\nConditions: Foggy\n\nSan Francisco is known for its microclimates and frequent fog, especially during the summer months. The temperature of 60°F (about 15.5°C) is quite typical for the city, which tends to have mild temperatures year-round. The fog, often referred to as "Karl the Fog" by locals, is a characteristic feature of San Francisco\'s weather, particularly in the mornings and evenings.\n\nIs there anything else you\'d like to know about the weather in San Francisco or any other location?"
```
Now when we pass the same <code>"thread_id"</code>, the conversation context is retained via the saved state (i.e. stored list of messages)
```python
final_state=app.invoke(
{"messages":[{"role":"user","content":"what about ny"}]},
config={"configurable":{"thread_id":42}}
)
final_state["messages"][-1].content
```
```
"Based on the search results, I can tell you that the current weather in New York City is:\n\nTemperature: 90 degrees Fahrenheit (approximately 32.2 degrees Celsius)\nConditions: Sunny\n\nThis weather is quite different from what we just saw in San Francisco. New York is experiencing much warmer temperatures right now. Here are a few points to note:\n\n1. The temperature of 90°F is quite hot, typical of summer weather in New York City.\n2. The sunny conditions suggest clear skies, which is great for outdoor activities but also means it might feel even hotter due to direct sunlight.\n3. This kind of weather in New York often comes with high humidity, which can make it feel even warmer than the actual temperature suggests.\n\nIt's interesting to see the stark contrast between San Francisco's mild, foggy weather and New York's hot, sunny conditions. This difference illustrates how varied weather can be across different parts of the United States, even on the same day.\n\nIs there anything else you'd like to know about the weather in New York or any other location?"
```
</details>
> [!TIP]
> LangGraph is a **low-level** framework that allows you to implement any custom agent
architectures. Click on the low-level implementation below to see how to implement a
# Note that we're (optionally) passing the memory when compiling the graph
app=workflow.compile(checkpointer=checkpointer)
# Use the Runnable
# Use the agent
final_state=app.invoke(
{"messages":[HumanMessage(content="what is the weather in sf")]},
{"messages":[{"role":"user","content":"what is the weather in sf"}]},
config={"configurable":{"thread_id":42}}
)
final_state["messages"][-1].content
```
```
'The current weather in San Francisco is as follows:\n- Temperature: 60.1°F (15.6°C)\n- Condition: Partly cloudy\n- Wind: 5.6 mph (9.0 kph) from SSW\n- Humidity: 83%\n- Visibility: 9.0 miles (16.0 km)\n- UV Index: 4.0\n\nFor more details, you can visit [Weather API](https://www.weatherapi.com/).'
```
<b>Step-by-step Breakdown</b>:
Now when we pass the same `"thread_id"`, the conversation context is retained via the saved state (i.e. stored list of messages)
<details>
<summary>Initialize the model and tools.</summary>
<ul>
<li>
We use <code>ChatAnthropic</code> as our LLM. <strong>NOTE:</strong> we need to make sure the model knows that it has these tools available to call. We can do this by converting the LangChain tools into the format for OpenAI tool calling using the <code>.bind_tools()</code> method.
</li>
<li>
We define the tools we want to use - a search tool in our case. It is really easy to create your own tools - see documentation here on how to do that <a href="https://python.langchain.com/docs/how_to/custom_tools/">here</a>.
</li>
</ul>
</details>
```python
final_state=app.invoke(
{"messages":[HumanMessage(content="what about ny")]},
config={"configurable":{"thread_id":42}}
)
final_state["messages"][-1].content
```
<details>
<summary>Initialize graph with state.</summary>
```
'The current weather in New York is as follows:\n- Temperature: 20.3°C (68.5°F)\n- Condition: Overcast\n- Wind: 2.2 mph from the north\n- Humidity: 65%\n- Cloud Cover: 100%\n- UV Index: 5.0\n\nFor more details, you can visit [Weather API](https://www.weatherapi.com/).'
```
<ul>
<li>We initialize graph (<code>StateGraph</code>) by passing state schema (in our case <code>MessagesState</code>)</li>
<li><code>MessagesState</code> is a prebuilt state schema that has one attribute -- a list of LangChain <code>Message</code> objects, as well as logic for merging the updates from each node into the state.</li>
</ul>
</details>
### Step-by-step Breakdown:
<details>
<summary>Define graph nodes.</summary>
1. <details>
<summary>Initialize the model and tools.</summary>
There are two main nodes we need:
- we use `ChatOpenAI` as our LLM. **NOTE:** we need make sure the model knows that it has these tools available to call. We can do this by converting the LangChain tools into the format for OpenAI tool calling using the `.bind_tools()` method.
- we define the tools we want to use -- a web search tool in our case. It is really easy to create your own tools - see documentation here on how to do that [here](https://python.langchain.com/docs/modules/agents/tools/custom_tools).
</details>
2. <details>
<summary>Initialize graph with state.</summary>
<ul>
<li>The <code>agent</code> node: responsible for deciding what (if any) actions to take.</li>
<li>The <code>tools</code> node that invokes tools: if the agent decides to take an action, this node will then execute that action.</li>
</ul>
</details>
- we initialize graph (`StateGraph`) by passing state schema (in our case `MessagesState`)
-`MessagesState` is a prebuilt state schema that has one attribute -- a list of LangChain `Message` objects, as well as logic for merging the updates from each node into the state
</details>
3. <details>
<summary>Define graph nodes.</summary>
<details>
<summary>Define entry point and graph edges.</summary>
There are two main nodes we need:
- The `agent` node: responsible for deciding what (if any) actions to take.
- The `tools` node that invokes tools: if the agent decides to take an action, this node will then execute that action.
</details>
4. <details>
<summary>Define entry point and graph edges.</summary>
First, we need to set the entry point for graph execution - <code>agent</code> node.
First, we need to set the entry point for graph execution - `agent` node.
Then we define one normal and one conditional edge. Conditional edge means that the destination depends on the contents of the graph's state (<code>MessagesState</code>). In our case, the destination is not known until the agent (LLM) decides.
Then we define one normal and one conditional edge. Conditional edge means that the destination depends on the contents of the graph's state (`MessageState`). In our case, the destination is not known until the agent (LLM) decides.
<ul>
<li>Conditional edge: after the agent is called, we should either:
<ul>
<li>a. Run tools if the agent said to take an action, OR</li>
<li>b. Finish (respond to the user) if the agent did not ask to run tools</li>
</ul>
</li>
<li>Normal edge: after the tools are invoked, the graph should always return to the agent to decide what to do next</li>
</ul>
</details>
- Conditional edge: after the agent is called, we should either:
- a. Run tools if the agent said to take an action, OR
- b. Finish (respond to the user) if the agent did not ask to run tools
- Normal edge: after the tools are invoked, the graph should always return to the agent to decide what to do next
</details>
5. <details>
<summary>Compile the graph.</summary>
<details>
<summary>Compile the graph.</summary>
- When we compile the graph, we turn it into a LangChain [Runnable](https://python.langchain.com/v0.2/docs/concepts/#runnable-interface), which automatically enables calling `.invoke()`, `.stream()` and `.batch()` with your inputs
- We can also optionally pass checkpointer object for persisting state between graph runs, and enabling memory, human-in-the-loop workflows, time travel and more. In our case we use `MemorySaver` - a simple in-memory checkpointer
</details>
6. <details>
<summary>Execute the graph.</summary>
<ul>
<li>
When we compile the graph, we turn it into a LangChain
which automatically enables calling <code>.invoke()</code>, <code>.stream()</code> and <code>.batch()</code>
with your inputs
</li>
<li>
We can also optionally pass checkpointer object for persisting state between graph runs, and enabling memory,
human-in-the-loop workflows, time travel and more. In our case we use <code>MemorySaver</code> -
a simple in-memory checkpointer
</li>
</ul>
</details>
1. LangGraph adds the input message to the internal state, then passes the state to the entrypoint node, `"agent"`.
2. The `"agent"` node executes, invoking the chat model.
3. The chat model returns an `AIMessage`. LangGraph adds this to the state.
4. Graph cycles the following steps until there are no more `tool_calls` on `AIMessage`:
- If `AIMessage` has `tool_calls`, `"tools"` node executes
- The `"agent"` node executes again and returns `AIMessage`
5. Execution progresses to the special `END` value and outputs the final state.
And as a result, we get a list of all our chat messages as output.
</details>
<details>
<summary>Execute the graph.</summary>
<ol>
<li>LangGraph adds the input message to the internal state, then passes the state to the entrypoint node, <code>"agent"</code>.</li>
<li>The <code>"agent"</code> node executes, invoking the chat model.</li>
<li>The chat model returns an <code>AIMessage</code>. LangGraph adds this to the state.</li>
<li>Graph cycles the following steps until there are no more <code>tool_calls</code> on <code>AIMessage</code>:
<ul>
<li>If <code>AIMessage</code> has <code>tool_calls</code>, <code>"tools"</code> node executes</li>
<li>The <code>"agent"</code> node executes again and returns <code>AIMessage</code></li>
</ul>
</li>
<li>Execution progresses to the special <code>END</code> value and outputs the final state. And as a result, we get a list of all our chat messages as output.</li>
</ol>
</details>
</details>
## Documentation
* [Tutorials](https://langchain-ai.github.io/langgraph/tutorials/): Learn to build with LangGraph through guided examples.
* [How-to Guides](https://langchain-ai.github.io/langgraph/how-tos/): Accomplish specific things within LangGraph, from streaming, to adding memory & persistence, to common design patterns (branching, subgraphs, etc.), these are the place to go if you want to copy and run a specific code snippet.
* [Conceptual Guides](https://langchain-ai.github.io/langgraph/concepts/): In-depth explanations of the key concepts and principles behind LangGraph, such as nodes, edges, state and more.
* [API Reference](https://langchain-ai.github.io/langgraph/reference/graphs/): Review important classes and methods, simple examples of how to use the graph and checkpointing APIs, higher-level prebuilt components and more.
* [Conceptual Guides](https://langchain-ai.github.io/langgraph/concepts/high_level/): In-depth explanations of the key concepts and principles behind LangGraph, such as nodes, edges, state and more.
* [API Reference](https://langchain-ai.github.io/langgraph/reference/graphs/): Review important classes and methods, simple examples of how to use the graph and checkpointing APIs, higher-level prebuilt components and more.
* [LangGraph Platform](https://langchain-ai.github.io/langgraph/concepts/#langgraph-platform): LangGraph Platform is a commercial solution for deploying agentic applications in production, built on the open-source LangGraph framework.
## Resources
* [Built with LangGraph](https://www.langchain.com/built-with-langgraph): Hear how industry leaders use LangGraph to ship powerful, production-ready AI applications.
## Contributing
For more information on how to contribute, see [here](https://github.com/langchain-ai/langgraph/blob/main/CONTRIBUTING.md).
Welcome to LangGraph, a Python library for building complex, scalable AI agents using graph-based state machines. In this guide, we'll explore the core concepts behind LangGraph and why it's uniquely suited for creating reliable, fault-tolerant agent systems. We assume you have already learned the basic covered in the [introduction tutorial](https://langchain-ai.github.io/langgraph/tutorials/introduction/#requirements) and want to deepen your understanding of LangGraph's underlying design and inner workings.
First off, why graphs?
## Background: Agents & AI Workflows as Graphs
While everyone has a slightly different definition of what constitutes an "AI Agent", we will take "agent" to mean any system that tasks a language model with controlling a looping workflow and takes actions. The prototypical LLM agent uses a ~["reasoning and action" (ReAct)](https://arxiv.org/abs/2210.03629)-style design, applying an LLM to power a basic loop with the following steps:
- reason and plan actions to take
- take actions using tools (regular software functions)
- observe the effects of the tools and re-plan or react as appropriate
While LLM agents are surprisingly effective at this, the naive agent loop doesn't deliver the [reliability users expect at scale](https://en.wikipedia.org/wiki/High_availability). They're beautifully stochastic. Well-designed systems take advantage of that randomness and apply it sensibly within a well-designed composite system and make that system **tolerant** to mistakes in the LLM's outputs, because mistakes **will** occur.
We think agents are exciting and new, but AI design patterns should apply applicable good engineering practices from Software 2.0. Some similarities include:
- AI applications must balance autonomous operations with user control.
- Agent applications resemble distributed systems in their need for error tolerance and correction.
- Multi-agent systems resemble multi-player web apps in their need for parallelism + conflict resolution.
- Everyone loves an undo button and version control.
LangGraph's primary [StateGraph](https://langchain-ai.github.io/langgraph/reference/graphs/#langgraph.graph.StateGraph) abstraction is designed to support these and other needs, providing an API that is lower level than other agent frameworks such as LangChain's [AgentExecutor](https://python.langchain.com/v0.1/docs/modules/agents/) to give you full control of where and how to apply "AI."
It extends Google's [Pregel](https://research.google/pubs/pregel-a-system-for-large-scale-graph-processing/) graph processing framework to provide fault tolerance and recovery when running long or error-prone workloads. When developing, you can focus on a local action or task-specific agent, and the system composes these actions to form a more capable and scalable application.
Its parallelism and `State` reduction functionality let you control what happens if, for example, multiple agents return conflicting information.
And finally, its persistent, versioned checkpointing system lets you roll back the agent's state, explore other paths, and maintain full control of what is going on.
The following sections go into greater detail about how and why all of this works.
## Core Design
At its core, LangGraph models agent workflows as state machines. You define the behavior of your agents using three key components:
1.`State`: A shared data structure that represents the current snapshot of your application. It can be any Python type, but is typically a `TypedDict` or Pydantic `BaseModel`.
2.`Nodes`: Python functions that encode the logic of your agents. They receive the current `State` as input, perform some computation or side-effect, and return an updated `State`.
3.`Edges`: Control flow rules that determine which `Node` to execute next based on the current `State`. They can be conditional branches or fixed transitions.
By composing `Nodes` and `Edges`, you can create complex, looping workflows that evolve the `State` over time. The real power, though, comes from how LangGraph manages that `State`.
Or in short: _nodes do the work. edges tell what to do next_.
LangGraph's underlying graph algorithm uses [message passing](https://en.wikipedia.org/wiki/Message_passing) to define a general program. When a `Node` completes, it sends a message along one or more edges to other node(s). These nodes run their functions, pass the resulting messages to the next set of nodes, and on and on it goes. Inspired by [Pregel](https://research.google/pubs/pregel-a-system-for-large-scale-graph-processing/), the program proceeds in discrete "super-steps" that are all executed conceptually in parallel. Whenever the graph is run, all the nodes start in an `inactive` state. Whenever an incoming edge (or "channel") receives a new message (state), the node becomes `active`, runs the function, and responds with updates. At the end of each superstep, each node votes to `halt` by marking itself as `inactive` if it has no more incoming messages. The graph terminates when all nodes are `inactive` and when no messages are in transit.
We will go through a full execution of a StateGraph later, but first, lets explore these concepts in more detail.
## Nodes
In StateGraph, nodes are typically python functions (sync or `async`) where the **first** positional argument is the [state](#state-management), and (optionally), the **second** positional argument is a "config", containing optional [configurable parameters](#configuration) (such as a `thread_id`).
Similar to `NetworkX`, you add these nodes to a graph using the [add_node](https://langchain-ai.github.io/langgraph/reference/graphs/#langgraph.graph.MessageGraph) method:
Behind the scenes, functions are converted to [RunnableLambda's](https://api.python.langchain.com/en/latest/runnables/langchain_core.runnables.base.RunnableLambda.html#langchain_core.runnables.base.RunnableLambda), which add batch and async support to your function, along with native tracing and debugging.
## Edges
Edges define how the logic is routed and how the graph decides to stop. Similar to nodes, they accept the current `state` of the graph and return a value.
By default, the value is the name of the node or nodes to send the state to next. All those nodes will be run in parallel as a part of the next superstep.
If you want to reuse an edge, you can optionally provide a dictionary that maps the edge's output to the name of the next node.
If you **always** want to go from node A to node B, you can use the [add_edge](https://langchain-ai.github.io/langgraph/reference/graphs/#langgraph.graph.StateGraph.add_edge) method directly.
If you want to **optionally** route to 1 or more edges (or optionally terminate), you can use the [add_conditional_edges](https://langchain-ai.github.io/langgraph/reference/graphs/#langgraph.graph.StateGraph.add_conditional_edges) method.
If a node has multiple out-going edges, **all** of those destination nodes will be executed in parallel as a part of the next superstep.
## State Management
LangGraph introduces two key ideas to state management: state schemas and reducers.
The state schema defines the type of the object that is given to each of the graph's `Node`.
Reducers define how to apply `Node` outputs to the current `State`. For example, you might use a reducer to merge a new dialogue response into a conversation history, or average together outputs from multiple agent nodes. By annotating your `State` fields with reducer functions, you can precisely control how data flows through your application.
We'll illustrate how reducers work with an example. Compare the following two `State`. Can you guess the output in both case?
In the first case (`StateA`), the result is "1", since the default **reducer** for your state is a direct overwrite.
In the second case (`StateB`), the result is "6" since we have have created the `add` function as the **reducer**. This function takes the existing state (for that field) and the state update (if provided) and returns the updated value for that state.
In general, **reducers** provided as annotations tell the graph **how to process updates for this field**.
While we typically use `TypedDict` as the graph's `state_schema` (i.e., `State`), it can be almost any [type](https://docs.python.org/3/library/stdtypes.html#type-objects), meaning the following graph is also completely valid:
```python
# Analogous to StateA above
builder=StateGraph(int)
builder.add_node("my_node",lambdastate:1)
builder.add_edge(START,"my_node")
builder.add_edge("my_node",END)
builder.compile().invoke(5)
# Analogous to StateB
defadd(left,right):
returnleft+right
builder=StateGraph(Annotated[int,add])
builder.add_node("my_node",lambdastate:1)
builder.add_edge(START,"my_node")
builder.add_edge("my_node",END)
graph=builder.compile()
graph.invoke(5)
```
This also means you can [use a Pydantic BaseModel](https://langchain-ai.github.io/langgraph/how-tos/state-model/) as your graph state to add **default values** and additional data validation.
When building simple chatbots like ChatGPT, the state can be as simple as a list of chat messages. This is the state used by [MessageGraph](https://langchain-ai.github.io/langgraph/reference/graphs/?h=message+graph#langgraph.graph.MessageGraph) (a light wrapper of `StateGraph`), which is only slightly more involved than the following:
```python
builder=StateGraph(Annotated[list,add])
```
Using a shared state within a graph comes with some design tradeoffs. For instance, you may think it feels like using dreaded global variables (though this can be addressed by namespacing arguments). However, sharing a typed state provides a number of benefits relevant to building AI workflows, including:
1. The data flow is fully inspectable before and after each "superstep".
2. The state is mutable, making it easy to let users or other software write to the same state between supersteps to control an agent's direction (using [update_state](https://langchain-ai.github.io/langgraph/reference/graphs/#langgraph.graph.graph.CompiledGraph.update_state)).
3. It is well-defined when checkpointing, making it easy to save and resume or even fully version control the execution of your entire workflows in whatever storage backend you wish.
We will talk about checkpointing more in the next section.
## Persistence
Any "intelligent" system needs memory to function. AI agents are no different, requiring memory across one or more timeframes:
- they _always_ need to remember the steps already taken **within this task** (to avoid repeating itself when answering a given query).
- they _typically_ need to remember the previous turns within a multi-turn conversation with a user (for coreference resolution and additional context).
- they _ideally_ need to "remember" context from previous interactions with the user and from actions in a given "environment" (such as an application context) to be more personalized and efficient in its behavior.
That last form of memory covers a lot (personalization, optimization, continual learning, etc.) and is beyond the scope of this conversation, although it can be easily integrated in any LangGraph workflow, and we are actively exploring the best way to expose this functionality natively.
The first two forms of memory are natively supported by the [StateGraph](https://langchain-ai.github.io/langgraph/reference/graphs/#langgraph.graph.StateGraph) API via [checkpointers](https://langchain-ai.github.io/langgraph/reference/checkpoints/#basecheckpointsaver).
#### Checkpoints
A checkpoint represents the state of a `thread` within a (potentially) multi-turn interaction between your application and a user (or users or other systems). Checkpoints that are made _within_ a single run will have a set of `next` nodes that will be executed when starting from this state. Checkpoints that are made at the end of a given run are identical, except there are no `next` nodes to transition to (the graph is awaiting user input).
Checkpointing supports chat memory and much more, letting you tag and persist every state your system has taken, regardless of whether it is within a single run or across many turns. Let's explore a bit why that is useful.
#### Single-turn Memory
**Within** a given run, each step of the agent is checkpointed. This means you could ask your agent to go create world peace. In the likely scenario that it runs into an error as it fails to do so, you can resume its quest at any time by resuming from one of its saved checkpoints.
This also lets you build **human-in-the-loop** workflows, common in use cases like [customer support bots](https://langchain-ai.github.io/langgraph/tutorials/customer-support/customer-support/), [programming assistants](https://langchain-ai.github.io/langgraph/tutorials/usaco/usaco/), and other applications. Before or after executing a given node, you can `interrupt` the graph's execution and "escalate" control to a user or support person. That person may respond immediately. Or they could respond a month from now. Either way, your workflow can resume at any time as if no time had passed at all.
#### Multi-turn Memory
Checkpoints are saved under a "thread_id" to support multi-turn interactions between users and your system. To the developer, there is absolutely no difference in how you configure your graph to add multi-turn memory support, since the checkpointing works the same throughout.
If you have some portion of state that you want to retain across turns and some state that you want to treat as "ephemeral", you can always clear the relevant state in the graph's final node.
Using checkpointing is as easy as calling `compile(checkpointer=my_checkpointer)` and then invoking it with a `thread_id` within its `configurable` parameters. You can see more in the following sections!
## Threads
Threads in LangGraph represent separate **sessions** of a graph. They organize state checkpoints within discrete sessions to facilitate multi-conversation and multi-user support in an application.
A typical chat bot application would have multiple threads for each user. Each thread represents a single conversation, with its own persistent chat history and other state. Checkpoints within a thread can be rewound and branched as needed.
Threads in LangGraph are distinct from [operating system threads](https://docs.python.org/3/library/threading.html), which are units of execution managed by the OS. They are more akin to a [conversational thread](<https://en.wikipedia.org/wiki/Thread_(online_communication)>) in email, twitter, and other messaging apps.
When a `StateGraph` is compiled with a checkpointer, each invocation of the graph requires a `thread_id` to be provided via [configuration (see below)](#configuration).
## Configuration
For any given graph deployment, you'll likely want some amount of configurable values that you can control at runtime. These differ from the graph **inputs** in that they aren't meant to be treated as state variables. They are more akin to "[out-of-band](https://en.wikipedia.org/wiki/Out-of-band)" communication.
A common example is a conversational `thread_id`, a `user_id`, a choice of which LLM to use, how many documents to return in a retriever, etc. While you **could** pass this within the state, it is nicer to separate out from the regular data flow. Configurable values are also automatically added to LangSmith traces as [metadata](https://docs.smith.langchain.com/concepts/tracing#metadata).
#### Example
Let's review another example to see how our multi-turn memory works! Can you guess what `result` and `result2` look like if you run this graph?
For the first run, no checkpoint existed, so the graph ran on the raw input. The "total" value is incremented from 1 to 2, and the "turn" is set to "First Turn".
For the second run, the user provides an update to "turn" but no total! Since we are loading from the state, the previous result is incremented by one (in our "add_one" node), and the "turn" is overwritten by the user.
For the third run, the "turn" remains the same, since it is loaded from the checkpoint but not overwritten by the user. The "total" is incremented by the value provided by the user, since this is **reduced** (i.e., used to update the existing value) by the `add` function.
For the fourth run, we are using a **new thread id** for which no checkpoint is found, so the result is just the user's provided **total** incremented by one.
You probably noticed that this user-facing behavior is equivalent to running the following **without a checkpointer**.
Run this for yourself to confirm equivalence. User inputs and checkpoint loading is treated more or less the same as any other **state update**.
Now that we've introduced the core concepts behind LangGraph, it may be instructive to walk through an end-to-end example to see how all the pieces fit together.
## Data flow of a single execution of a StateGraph
As engineers, we are never really satisfied until we know what's going on "under the hood". In the previous sections, we explained some of the LangGraph's core concepts. Now it's time to really show how they fit together.
Let's extend our toy example above with a conditional edge and then walk through two consecutive invocations.
To inspect the trace of this run, check out the [LangSmith link here](https://smith.langchain.com/public/0c543370-d459-4b8d-9962-058f67bdc9ce/r). We'll walk through the execution below:
1. First, the graph looks for a checkpoint. None is found, so the state is thus initialized with a total of 0.
2. Next, the graph applies the user's input as an update to the state. The reducer adds the input (1) to the existing value (0). At the end of this superstep, the total is (1).
3. After that, the "add_one" node is called, returning 1.
4. Next, the reducer adds this update to the existing total (1). The state is now 2.
5. Then, the conditional edge "`route`" is called. Since the value is less than 6, we continue to the 'double' node.
6. Double takes the existing state (2), and returns it. The reducer is then called and adds it to the existing state. The state is now 4.
7. The graph then loops back through add_one (5), checks the conditional edge and proceeds to since it's < 6. After doubling, the total is (10).
8. The fixed edge loops back to add_one (11), checks the conditional edge, and since it is greater than 6, the program terminates.
For our second run, we will use the same configuration:
To inspect the trace of this run, check out the [LangSmith link here](https://smith.langchain.com/public/494f1817-46f5-4051-b41c-2dc416ce8b4d/r). We'll walk through the execution below:
1. First, it applies the update from the user's input. The `add`**reducer** updates the total from 0 to -2.
2. Next, the graph looks for the checkpoint. It loads it to memory as the initial state. Total is (9) now ((-2) + 11).
3. After that, the 'add_one' node is called with this state. It returns 10.
4. That update is applied using the reducer, raising the value to 10.
5. Next, the "route" conditional edge is triggered. Since the value is greater than 6, we terminate the program, ending where we started at (11).
Welcome to the LangGraph how-to guides! These guides provide practical, step-by-step instructions for accomplishing key tasks in LangGraph.
## Core
The core guides show how to address common needs when building out AI workflows, with special focus placed on [ReAct](https://arxiv.org/abs/2210.03629)-style agents with [tool calling](https://python.langchain.com/docs/modules/model_io/chat/function_calling/).
- [ReAct agent](create-react-agent.ipynb): How to create a tool-calling agent that **Re**asons and **Act**s to accomplish tasks
- [Persistence](persistence.ipynb): How to give your graph "memory" and resilience by saving and loading state
- [Time travel](time-travel.ipynb): How to navigate and manipulate graph state history once it's persisted
- [Async execution](async.ipynb): How to run nodes asynchronously for improved performance
- [Streaming responses](streaming-tokens.ipynb): How to stream agent responses in real-time
- [Visualization](visualization.ipynb): How to visualize your graphs
- [Configuration](configuration.ipynb): How to indicate that a graph can swap out configurable components
### Design patterns
Recipes showing how to apply common design patterns in your workflows:
- [Subgraphs](subgraph.ipynb): How to compose subgraphs within a larger graph
- [Branching](branching.ipynb): How to create branching logic in your graphs for parallel node execution
- [Map-reduce](map-reduce.ipynb): How to branch **different views** of the state for parallel node execution (even applying the same node in parallel N times)
- [Human-in-the-loop](human-in-the-loop.ipynb): How to incorporate human feedback and intervention
The following examples are useful especially if you are used to LangChain's AgentExecutor configurations.
- [Force calling a tool first](force-calling-a-tool-first.ipynb): Define a fixed workflow before ceding control to the ReAct agent
- [Pass run time values to tools](pass-run-time-values-to-tools.ipynb): Pass values that are only known at run time to tools (e.g., the ID of the user who made the request)
- [Dynamic direct return](dynamically-returning-directly.ipynb): Let the LLM decide whether the graph should finish after a tool is run or whether the LLM should be able to review the output and keep going
- [Respond in structured format](respond-in-format.ipynb): Let the LLM use tools or populate schema to provide the user. Useful if your agent should generate structured content
- [Managing agent steps](managing-agent-steps.ipynb): How to format the intermediate steps of your workflow for the agent
### Alternative ways to define state
- [Pydantic state](state-model.ipynb): Use a Pydantic model as your state
### Structured output
- [Extraction with re-prompting](./extraction/retries.ipynb): How to generate complex nested schemas using JSONPatch retries, for when function calling is insufficient, and regular reprompting still fails to generate valid results
- The `MessageGraph` contains the agent's "Memory"
- Conditional edges enable dynamic routing between the chatbot, tools, and the user
- Persistence makes it easy to stop, resume, and even rewind for full control over your application
With LangGraph, you can build complex, stateful agents without getting bogged down in manual state and interrupt management. Just define your nodes, edges, and state schema - and let the graph take care of the rest.
## Tutorials
Consult the [Tutorials](tutorials/index.md) to learn more about building with LangGraph, including advanced use cases.
## How-To Guides
Check out the [How-To Guides](how-tos/index.md) for instructions on handling common tasks with LangGraph
## Reference
For documentation on the core APIs, check out the [Reference](reference/graphs.md) docs.
## Conceptual Guides
Once you've learned the basics, if you want to further understand LangGraph's core abstractions, check out the [Conceptual Guides](./concepts/index.md).
## Why LangGraph?
LangGraph is framework agnostic (each node is a regular python function). It extends the core Runnable API (shared interface for streaming, async, and batch calls) to make it easy to:
- Seamless state management across multiple turns of conversation or tool usage
- The ability to flexibly route between nodes based on dynamic criteria
- Smooth switching between LLMs and human intervention
- Persistence for long-running, multi-session applications
If you're building a straightforward DAG, Runnables are a great fit. But for more complex, stateful applications with nonlinear flows, LangGraph is the perfect tool for the job.
You can [compile](https://langchain-ai.github.io/langgraph/reference/graphs/#langgraph.graph.MessageGraph.compile) any LangGraph workflow with a [CheckPointer](https://langchain-ai.github.io/langgraph/reference/checkpoints/#basecheckpointsaver) to give your agent "memory" by persisting its state. This permits things like:
- Remembering things across multiple interactions
- Interrupting to wait for user input
- Resilience for long-running, error-prone agents
- Time travel retry and branch from a previous checkpoint
### Checkpoint
::: langgraph.checkpoint.Checkpoint
### BaseCheckpointSaver
::: langgraph.checkpoint.base.BaseCheckpointSaver
handler: python
### SerializerProtocol
::: langgraph.checkpoint.SerializerProtocol
handler: python
## Implementations
LangGraph also natively provides the following checkpoint implementations.
Graphs are the core abstraction of LangGraph. Each [StateGraph](#stategraph) implementation is used to create graph workflows. Once compiled, you can run the [CompiledGraph](#compiledgraph) to run the application.
## StateGraph
```python
fromlanggraph.graphimportStateGraph
fromtyping_extensionsimportTypedDict
classMyState(TypedDict)
...
graph=StateGraph(MyState)
```
::: langgraph.graph.StateGraph
handler: python
## MessageGraph
::: langgraph.graph.message.MessageGraph
## CompiledGraph
::: langgraph.graph.graph.CompiledGraph
handler: python
## Constants
The following constants and classes are used to help control graph execution.
## START
START is a string constant (`"__start__"`) that serves as a "virtual" node in the graph.
Adding an edge (or conditional edges) from `START` to node one or more nodes in your graph
will direct the graph to begin execution there.
```python
fromlanggraph.graphimportSTART
...
builder.add_edge(START,"my_node")
# Or to add a conditional starting point
builder.add_conditional_edges(START,my_condition)
```
## END
END is a string constant (`"__end__"`) that serves as a "virtual" node in the graph. Adding
an edge (or conditional edges) from one or more nodes in your graph to the `END` "node" will
direct the graph to cease execution as soon as it reaches this point.
```python
fromlanggraph.graphimportEND
...
builder.add_edge("my_node",END)# Stop any time my_node completes
Welcome to the LangGraph Tutorials! These notebooks introduce LangGraph through building various language agents and applications.
## Introduction to LangGraph
Learn the basics of LangGraph through the onboarding tutorials.
- [Introduction to LangGraph](introduction.ipynb)
## Use cases
Learn from example implementations of graphs designed for specific scenarios and that implement common design patterns.
#### Chatbots
- [Customer Support](customer-support/customer-support.ipynb): Build a customer support chatbot to manage flights, hotel reservations, car rentals, and other tasks
- [Info Gathering](chatbots/information-gather-prompting.ipynb): Build an information gathering chatbot
- [Code Assistant](code_assistant/langgraph_code_assistant.ipynb): Building a code analysis and generation assistant
#### Multi-Agent Systems
- [Collaboration](multi_agent/multi-agent-collaboration.ipynb): Enabling two agents to collaborate on a task
- [Supervision](multi_agent/agent_supervisor.ipynb): Using an LLM to orchestrate and delegate to individual agents
- [Hierarchical Teams](multi_agent/hierarchical_agent_teams.ipynb): Orchestrating nested teams of agents to solve problems
- [Corrective RAG with local models](rag/langgraph_crag_local.ipynb)
- [Self-RAG](rag/langgraph_self_rag.ipynb)
- [Self-RAG with local models](rag/langgraph_self_rag_local.ipynb)
- [Web Research (STORM)](storm/storm.ipynb): Generating Wikipedia-like articles via research and multi-perspective QA
#### Planning Agents
- [Plan-and-Execute](plan-and-execute/plan-and-execute.ipynb): Implementing a basic planning and execution agent
- [Reasoning without Observation](rewoo/rewoo.ipynb): Reducing re-planning by saving observations as variables
- [LLMCompiler](llm-compiler/LLMCompiler.ipynb): Streaming and eagerly executing a DAG of tasks from a planner
#### Reflection & Critique
- [Basic Reflection](reflection/reflection.ipynb): Prompting the agent to reflect on and revise its outputs
- [Reflexion](reflexion/reflexion.ipynb): Critiquing missing and superfluous details to guide next steps
- [Language Agent Tree Search](lats/lats.ipynb): Using reflection and rewards to drive a tree search over agents
- [Self-Discovering Agent](self-discover/self-discover.ipynb): Analyzing an agent that learns about its own capabilities
#### Evaluation
- [Agent-based](chatbot-simulation-evaluation/agent-simulation-evaluation.ipynb): Evaluating chatbots via simulated user interactions
- [Within LangSmith](chatbot-simulation-evaluation/langsmith-agent-simulation-evaluation.ipynb): Evaluating chatbots in LangSmith over a dialog dataset
#### Text Mining
- [TNT-LLM](tnt-llm/tnt-llm.ipynb): learn to build rich, interpretable taxonomies of user intentand using the classification system developed by Microsoft for their Bing Copilot application.
#### Competitive Programming
- [Can Language Models Solve Olympiad Programming?](usaco/usaco.ipynb): Build an agent with few-shot "episodic memory" and human-in-the-loop collaboration to solve problems from the USA Computing Olympiad; adapted from the [paper of the same name](https://arxiv.org/abs/2404.10952v1) by Shi, Tang, Narasimhan, and Yao.
#### Other Experimental Architectures
- [Web Navigation](web-navigation/web_voyager.ipynb): Building an agent that can navigate and interact with websites
"os.environ[\"LANGCHAIN_API_KEY\"] = getpass.getpass(\"LangSmith API Key:\")"
]
},
{
"cell_type": "markdown",
"id": "5dace4a9-7c9e-4da2-bf7b-e58d0d05581e",
"metadata": {},
"source": [
"## Create the LangChain agent\n",
"\n",
"First, we will create the LangChain agent. For more information on LangChain agents, see [this documentation](https://python.langchain.com/v0.2/docs/concepts/#agents)"
"We now define the graph state. The state for the traditional LangChain agent has a few attributes:\n",
"\n",
"1. `input`: This is the input string representing the main ask from the user, passed in as input.\n",
"2. `chat_history`: This is any previous conversation messages, also passed in as input.\n",
"3. `intermediate_steps`: This is list of actions and corresponding observations that the agent takes over time. This is updated each iteration of the agent.\n",
"4. `agent_outcome`: This is the response from the agent, either an AgentAction or AgentFinish. The AgentExecutor should finish when this is an AgentFinish, otherwise it should call the requested tools.\n"
"We now need to define a few different nodes in our graph.\n",
"In `langgraph`, a node can be either a function or a [runnable](https://python.langchain.com/v0.2/docs/concepts/#langchain-expression-language-lcel).\n",
"There are two main nodes we need for this:\n",
"\n",
"1. The agent: responsible for deciding what (if any) actions to take.\n",
"2. A function to invoke tools: if the agent decides to take an action, this node will then execute that action.\n",
"\n",
"We will also need to define some edges.\n",
"Some of these edges may be conditional.\n",
"The reason they are conditional is that based on the output of a node, one of several paths may be taken.\n",
"The path that is taken is not known until that node is run (the LLM decides).\n",
"\n",
"1. Conditional Edge: after the agent is called, we should either:\n",
" a. If the agent said to take an action, then the function to invoke tools should be called\n",
" b. If the agent said that it was finished, then it should finish\n",
"2. Normal Edge: after the tools are invoked, it should always go back to the agent to decide what to do next\n",
"\n",
"Let's define the nodes, as well as a function to decide how what conditional edge to take."
"# Define logic that will be used to determine which conditional edge to go down\n",
"def should_continue(data):\n",
" # If the agent outcome is an AgentFinish, then we return `exit` string\n",
" # This will be used when setting up the graph to define the flow\n",
" if isinstance(data[\"agent_outcome\"], AgentFinish):\n",
" return \"end\"\n",
" # Otherwise, an AgentAction is returned\n",
" # Here we return `continue` string\n",
" # This will be used when setting up the graph to define the flow\n",
" else:\n",
" return \"continue\""
]
},
{
"cell_type": "markdown",
"id": "c0b211f4-0c5c-4792-b18d-cd70907c71e7",
"metadata": {},
"source": [
"## Define the graph\n",
"\n",
"We can now put it all together and define the graph!"
]
},
{
"cell_type": "code",
"execution_count": 12,
"id": "c4054dde-4618-49b7-998a-daa0c1d6d6c0",
"metadata": {},
"outputs": [],
"source": [
"from langgraph.graph import END, StateGraph\n",
"\n",
"# Define a new graph\n",
"workflow = StateGraph(AgentState)\n",
"\n",
"# Define the two nodes we will cycle between\n",
"workflow.add_node(\"agent\", run_agent)\n",
"workflow.add_node(\"action\", execute_tools)\n",
"\n",
"# Set the entrypoint as `agent`\n",
"# This means that this node is the first one called\n",
"workflow.set_entry_point(\"agent\")\n",
"\n",
"# We now add a conditional edge\n",
"workflow.add_conditional_edges(\n",
" # First, we define the start node. We use `agent`.\n",
" # This means these are the edges taken after the `agent` node is called.\n",
" \"agent\",\n",
" # Next, we pass in the function that will determine which node is called next.\n",
" should_continue,\n",
" # Finally we pass in a mapping.\n",
" # The keys are strings, and the values are other nodes.\n",
" # END is a special node marking that the graph should finish.\n",
" # What will happen is we will call `should_continue`, and then the output of that\n",
" # will be matched against the keys in this mapping.\n",
" # Based on which one it matches, that node will then be called.\n",
" {\n",
" # If `tools`, then we call the tool node.\n",
" \"continue\": \"action\",\n",
" # Otherwise we finish.\n",
" \"end\": END,\n",
" },\n",
")\n",
"\n",
"# We now add a normal edge from `tools` to `agent`.\n",
"# This means that after `tools` is called, `agent` node is called next.\n",
"workflow.add_edge(\"action\", \"agent\")\n",
"\n",
"# Finally, we compile it!\n",
"# This compiles it into a LangChain Runnable,\n",
"# meaning you can use it as you would any other runnable\n",
"app = workflow.compile()"
]
},
{
"cell_type": "code",
"execution_count": 13,
"id": "214ae46e-c297-465d-86db-2b0312ed3530",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'agent_outcome': AgentActionMessageLog(tool='tavily_search_results_json', tool_input={'query': 'weather in San Francisco'}, log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\", message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})])}\n",
"----\n",
"{'intermediate_steps': [(AgentActionMessageLog(tool='tavily_search_results_json', tool_input={'query': 'weather in San Francisco'}, log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\", message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})]), \"[{'url': 'https://www.whereandwhen.net/when/north-america/california/san-francisco-ca/january/', 'content': 'Best time to go to San Francisco? Weather in San Francisco in january 2024 How was the weather last january? Here is the day by day recorded weather in San Francisco in january 2023: Seasonal average climate and temperature of San Francisco in january 8% 46% 29% 12% 8% Evolution of daily average temperature and precipitation in San Francisco in januaryWeather in San Francisco in january 2024. The weather in San Francisco in january comes from statistical data on the past years. You can view the weather statistics the entire month, but also by using the tabs for the beginning, the middle and the end of the month. ... 16-01-2023 45°F to 52°F. 17-01-2023 45°F to 54°F. 18-01-2023 47°F to ...'}]\")]}\n",
"----\n",
"{'agent_outcome': AgentFinish(return_values={'output': 'I found some information about the weather in San Francisco in January 2024, but it seems that the search results are not specific to the current weather. Would you like me to try a different search method to get the current weather in San Francisco?'}, log='I found some information about the weather in San Francisco in January 2024, but it seems that the search results are not specific to the current weather. Would you like me to try a different search method to get the current weather in San Francisco?')}\n",
"----\n",
"{'input': 'what is the weather in sf', 'chat_history': [], 'agent_outcome': AgentFinish(return_values={'output': 'I found some information about the weather in San Francisco in January 2024, but it seems that the search results are not specific to the current weather. Would you like me to try a different search method to get the current weather in San Francisco?'}, log='I found some information about the weather in San Francisco in January 2024, but it seems that the search results are not specific to the current weather. Would you like me to try a different search method to get the current weather in San Francisco?'), 'intermediate_steps': [(AgentActionMessageLog(tool='tavily_search_results_json', tool_input={'query': 'weather in San Francisco'}, log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\", message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})]), \"[{'url': 'https://www.whereandwhen.net/when/north-america/california/san-francisco-ca/january/', 'content': 'Best time to go to San Francisco? Weather in San Francisco in january 2024 How was the weather last january? Here is the day by day recorded weather in San Francisco in january 2023: Seasonal average climate and temperature of San Francisco in january 8% 46% 29% 12% 8% Evolution of daily average temperature and precipitation in San Francisco in januaryWeather in San Francisco in january 2024. The weather in San Francisco in january comes from statistical data on the past years. You can view the weather statistics the entire month, but also by using the tabs for the beginning, the middle and the end of the month. ... 16-01-2023 45°F to 52°F. 17-01-2023 45°F to 54°F. 18-01-2023 47°F to ...'}]\")]}\n",
"----\n"
]
}
],
"source": [
"inputs = {\"input\": \"what is the weather in sf\", \"chat_history\": []}\n",
"In this notebook we will create an agent with a search tool. However, at the start we will force the agent to call the search tool (and then let it do whatever it wants after). This is useful when you want to force agents to call particular tools, but still want flexibility of what happens after that.\n",
"\n",
"This examples builds off the base agent executor. It is highly recommended you learn about that executor before going through this notebook. You can find documentation for that example [here](./base.ipynb).\n",
"\n",
"Any modifications of that example are called below with **MODIFICATION**, so if you are looking for the differences you can just search for that."
"os.environ[\"LANGCHAIN_API_KEY\"] = getpass.getpass(\"LangSmith API Key:\")"
]
},
{
"cell_type": "markdown",
"id": "5dace4a9-7c9e-4da2-bf7b-e58d0d05581e",
"metadata": {},
"source": [
"## Create the LangChain agent\n",
"\n",
"First, we will create the LangChain agent. For more information on LangChain agents, see [this documentation](https://python.langchain.com/v0.2/docs/concepts/#agents)"
"We now define the graph state. The state for the traditional LangChain agent has a few attributes:\n",
"\n",
"1. `input`: This is the input string representing the main ask from the user, passed in as input.\n",
"2. `chat_history`: This is any previous conversation messages, also passed in as input.\n",
"3. `intermediate_steps`: This is list of actions and corresponding observations that the agent takes over time. This is updated each iteration of the agent.\n",
"4. `agent_outcome`: This is the response from the agent, either an AgentAction or AgentFinish. The AgentExecutor should finish when this is an AgentFinish, otherwise it should call the requested tools.\n"
"We now need to define a few different nodes in our graph.\n",
"In `langgraph`, a node can be either a function or a [runnable](https://python.langchain.com/v0.2/docs/concepts/#langchain-expression-language-lcel).\n",
"There are two main nodes we need for this:\n",
"\n",
"1. The agent: responsible for deciding what (if any) actions to take.\n",
"2. A function to invoke tools: if the agent decides to take an action, this node will then execute that action.\n",
"\n",
"We will also need to define some edges.\n",
"Some of these edges may be conditional.\n",
"The reason they are conditional is that based on the output of a node, one of several paths may be taken.\n",
"The path that is taken is not known until that node is run (the LLM decides).\n",
"\n",
"1. Conditional Edge: after the agent is called, we should either:\n",
" a. If the agent said to take an action, then the function to invoke tools should be called\n",
" b. If the agent said that it was finished, then it should finish\n",
"2. Normal Edge: after the tools are invoked, it should always go back to the agent to decide what to do next\n",
"\n",
"Let's define the nodes, as well as a function to decide how what conditional edge to take."
"# This compiles it into a LangChain Runnable,\n",
"# meaning you can use it as you would any other runnable\n",
"app = workflow.compile()"
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "214ae46e-c297-465d-86db-2b0312ed3530",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'agent_outcome': AgentActionMessageLog(tool='tavily_search_results_json', tool_input='what is the weather in sf', log='', message_log=[])}\n",
"----\n",
"{'intermediate_steps': [(AgentActionMessageLog(tool='tavily_search_results_json', tool_input='what is the weather in sf', log='', message_log=[]), \"[{'url': 'https://www.whereandwhen.net/when/north-america/california/san-francisco-ca/january/', 'content': 'Best time to go to San Francisco? Weather in San Francisco in january 2024 How was the weather last january? Here is the day by day recorded weather in San Francisco in january 2023: Seasonal average climate and temperature of San Francisco in january The climate of San Francisco in january is tolerableWeather in San Francisco in january 2024. The weather in San Francisco in january comes from statistical data on the past years. You can view the weather statistics the entire month, but also by using the tabs for the beginning, the middle and the end of the month. ... 15-01-2023 50°F to 52°F. 16-01-2023 45°F to 52°F. 17-01-2023 45°F to ...'}]\")]}\n",
"----\n",
"{'agent_outcome': AgentFinish(return_values={'output': 'The weather in San Francisco in January is typically tolerable, with temperatures ranging from 45°F to 52°F. If you need more specific and up-to-date information about the current weather in San Francisco, I can look it up for you.'}, log='The weather in San Francisco in January is typically tolerable, with temperatures ranging from 45°F to 52°F. If you need more specific and up-to-date information about the current weather in San Francisco, I can look it up for you.')}\n",
"----\n",
"{'input': 'what is the weather in sf', 'chat_history': [], 'agent_outcome': AgentFinish(return_values={'output': 'The weather in San Francisco in January is typically tolerable, with temperatures ranging from 45°F to 52°F. If you need more specific and up-to-date information about the current weather in San Francisco, I can look it up for you.'}, log='The weather in San Francisco in January is typically tolerable, with temperatures ranging from 45°F to 52°F. If you need more specific and up-to-date information about the current weather in San Francisco, I can look it up for you.'), 'intermediate_steps': [(AgentActionMessageLog(tool='tavily_search_results_json', tool_input='what is the weather in sf', log='', message_log=[]), \"[{'url': 'https://www.whereandwhen.net/when/north-america/california/san-francisco-ca/january/', 'content': 'Best time to go to San Francisco? Weather in San Francisco in january 2024 How was the weather last january? Here is the day by day recorded weather in San Francisco in january 2023: Seasonal average climate and temperature of San Francisco in january The climate of San Francisco in january is tolerableWeather in San Francisco in january 2024. The weather in San Francisco in january comes from statistical data on the past years. You can view the weather statistics the entire month, but also by using the tabs for the beginning, the middle and the end of the month. ... 15-01-2023 50°F to 52°F. 16-01-2023 45°F to 52°F. 17-01-2023 45°F to ...'}]\")]}\n",
"----\n"
]
}
],
"source": [
"inputs = {\"input\": \"what is the weather in sf\", \"chat_history\": []}\n",
"The `create_agent_executor` function is deprecated in favor of [create_react_agent](../chat_agent_executor_with_function_calling/high-level-tools.ipynb).\n",
"This was done to better align with the underlying model providers' migration from \"function calling\" to \"tool calling\", which typically supports parallel tool usage."
"In this notebook we will go over how to add a human-in-the-loop workflow to the base agent executor. We will use the human to approve\n",
"\n",
"This examples builds off the base agent executor. It is highly recommended you learn about that executor before going through this notebook. You can find documentation for that example [here](./base.ipynb).\n",
"\n",
"Any modifications of that example are called below with **MODIFICATION**, so if you are looking for the differences you can just search for that."
"os.environ[\"LANGCHAIN_API_KEY\"] = getpass.getpass(\"LangSmith API Key:\")"
]
},
{
"cell_type": "markdown",
"id": "5dace4a9-7c9e-4da2-bf7b-e58d0d05581e",
"metadata": {},
"source": [
"## Create the LangChain agent\n",
"\n",
"First, we will create the LangChain agent. For more information on LangChain agents, see [this documentation](https://python.langchain.com/v0.2/docs/concepts/#agents)"
"We now define the graph state. The state for the traditional LangChain agent has a few attributes:\n",
"\n",
"1. `input`: This is the input string representing the main ask from the user, passed in as input.\n",
"2. `chat_history`: This is any previous conversation messages, also passed in as input.\n",
"3. `intermediate_steps`: This is list of actions and corresponding observations that the agent takes over time. This is updated each iteration of the agent.\n",
"4. `agent_outcome`: This is the response from the agent, either an AgentAction or AgentFinish. The AgentExecutor should finish when this is an AgentFinish, otherwise it should call the requested tools.\n"
"We now need to define a few different nodes in our graph.\n",
"In `langgraph`, a node can be either a function or a [runnable](https://python.langchain.com/v0.2/docs/concepts/#langchain-expression-language-lcel).\n",
"There are two main nodes we need for this:\n",
"\n",
"1. The agent: responsible for deciding what (if any) actions to take.\n",
"2. A function to invoke tools: if the agent decides to take an action, this node will then execute that action.\n",
"\n",
"We will also need to define some edges.\n",
"Some of these edges may be conditional.\n",
"The reason they are conditional is that based on the output of a node, one of several paths may be taken.\n",
"The path that is taken is not known until that node is run (the LLM decides).\n",
"\n",
"1. Conditional Edge: after the agent is called, we should either:\n",
" a. If the agent said to take an action, then the function to invoke tools should be called\n",
" b. If the agent said that it was finished, then it should finish\n",
"2. Normal Edge: after the tools are invoked, it should always go back to the agent to decide what to do next\n",
"\n",
"Let's define the nodes, as well as a function to decide how what conditional edge to take."
"# This a helper class we have that is useful for running tools\n",
"# It takes in an agent action and calls that tool and returns the result\n",
"tool_executor = ToolExecutor(tools)\n",
"\n",
"\n",
"# Define the agent\n",
"def run_agent(data):\n",
" agent_outcome = agent_runnable.invoke(data)\n",
" return {\"agent_outcome\": agent_outcome}"
]
},
{
"cell_type": "markdown",
"id": "35ace508-d5fe-4139-a0f8-887e38047401",
"metadata": {},
"source": [
"**MODIFICATION**\n",
"\n",
"We modify the function that is calling the tool to first ask for user approval to continue. Note that this is a simple example and we could modify it to change the tool input, use some other channel besides input, etc."
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "2fecf5e0-9604-4992-9c82-b9627466cd32",
"metadata": {},
"outputs": [],
"source": [
"# Define the function to execute tools\n",
"def execute_tools(data):\n",
" # Get the most recent agent_outcome - this is the key added in the `agent` above\n",
"# Define logic that will be used to determine which conditional edge to go down\n",
"def should_continue(data):\n",
" # If the agent outcome is an AgentFinish, then we return `exit` string\n",
" # This will be used when setting up the graph to define the flow\n",
" if isinstance(data[\"agent_outcome\"], AgentFinish):\n",
" return \"end\"\n",
" # Otherwise, an AgentAction is returned\n",
" # Here we return `continue` string\n",
" # This will be used when setting up the graph to define the flow\n",
" else:\n",
" return \"continue\""
]
},
{
"cell_type": "markdown",
"id": "c0b211f4-0c5c-4792-b18d-cd70907c71e7",
"metadata": {},
"source": [
"## Define the graph\n",
"\n",
"We can now put it all together and define the graph!"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "c4054dde-4618-49b7-998a-daa0c1d6d6c0",
"metadata": {},
"outputs": [],
"source": [
"from langgraph.graph import END, StateGraph\n",
"\n",
"# Define a new graph\n",
"workflow = StateGraph(AgentState)\n",
"\n",
"# Define the two nodes we will cycle between\n",
"workflow.add_node(\"agent\", run_agent)\n",
"workflow.add_node(\"action\", execute_tools)\n",
"\n",
"# Set the entrypoint as `agent`\n",
"# This means that this node is the first one called\n",
"workflow.set_entry_point(\"agent\")\n",
"\n",
"# We now add a conditional edge\n",
"workflow.add_conditional_edges(\n",
" # First, we define the start node. We use `agent`.\n",
" # This means these are the edges taken after the `agent` node is called.\n",
" \"agent\",\n",
" # Next, we pass in the function that will determine which node is called next.\n",
" should_continue,\n",
" # Finally we pass in a mapping.\n",
" # The keys are strings, and the values are other nodes.\n",
" # END is a special node marking that the graph should finish.\n",
" # What will happen is we will call `should_continue`, and then the output of that\n",
" # will be matched against the keys in this mapping.\n",
" # Based on which one it matches, that node will then be called.\n",
" {\n",
" # If `tools`, then we call the tool node.\n",
" \"continue\": \"action\",\n",
" # Otherwise we finish.\n",
" \"end\": END,\n",
" },\n",
")\n",
"\n",
"# We now add a normal edge from `tools` to `agent`.\n",
"# This means that after `tools` is called, `agent` node is called next.\n",
"workflow.add_edge(\"action\", \"agent\")\n",
"\n",
"# Finally, we compile it!\n",
"# This compiles it into a LangChain Runnable,\n",
"# meaning you can use it as you would any other runnable\n",
"app = workflow.compile()"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "214ae46e-c297-465d-86db-2b0312ed3530",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'agent_outcome': AgentActionMessageLog(tool='tavily_search_results_json', tool_input={'query': 'weather in San Francisco'}, log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\", message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})])}\n",
"----\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"[y/n] continue with: tool='tavily_search_results_json' tool_input={'query': 'weather in San Francisco'} log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\" message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})]? y\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'intermediate_steps': [(AgentActionMessageLog(tool='tavily_search_results_json', tool_input={'query': 'weather in San Francisco'}, log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\", message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})]), \"[{'url': 'https://weatherspark.com/h/m/557/2024/1/Historical-Weather-in-January-2024-in-San-Francisco-California-United-States', 'content': 'January 2024 Weather History in San Francisco California, United States Daily Precipitation in January 2024 in San Francisco Observed Weather in January 2024 in San Francisco San Francisco Temperature History January 2024 Hourly Temperature in January 2024 in San Francisco Hours of Daylight and Twilight in January 2024 in San FranciscoThis report shows the past weather for San Francisco, providing a weather history for January 2024. It features all historical weather data series we have available, including the San Francisco temperature history for January 2024. You can drill down from year to month and even day level reports by clicking on the graphs.'}]\")]}\n",
"----\n",
"{'agent_outcome': AgentFinish(return_values={'output': \"It seems that I didn't find the current weather information for San Francisco. I recommend checking a reliable weather website or using a weather app to get the most up-to-date information.\"}, log=\"It seems that I didn't find the current weather information for San Francisco. I recommend checking a reliable weather website or using a weather app to get the most up-to-date information.\")}\n",
"----\n",
"{'input': 'what is the weather in sf', 'chat_history': [], 'agent_outcome': AgentFinish(return_values={'output': \"It seems that I didn't find the current weather information for San Francisco. I recommend checking a reliable weather website or using a weather app to get the most up-to-date information.\"}, log=\"It seems that I didn't find the current weather information for San Francisco. I recommend checking a reliable weather website or using a weather app to get the most up-to-date information.\"), 'intermediate_steps': [(AgentActionMessageLog(tool='tavily_search_results_json', tool_input={'query': 'weather in San Francisco'}, log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\", message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})]), \"[{'url': 'https://weatherspark.com/h/m/557/2024/1/Historical-Weather-in-January-2024-in-San-Francisco-California-United-States', 'content': 'January 2024 Weather History in San Francisco California, United States Daily Precipitation in January 2024 in San Francisco Observed Weather in January 2024 in San Francisco San Francisco Temperature History January 2024 Hourly Temperature in January 2024 in San Francisco Hours of Daylight and Twilight in January 2024 in San FranciscoThis report shows the past weather for San Francisco, providing a weather history for January 2024. It features all historical weather data series we have available, including the San Francisco temperature history for January 2024. You can drill down from year to month and even day level reports by clicking on the graphs.'}]\")]}\n",
"----\n"
]
}
],
"source": [
"inputs = {\"input\": \"what is the weather in sf\", \"chat_history\": []}\n",
"In this notebook we will go over how to build a basic agent executor where we custom handle how to manage the intermediate steps. Normally, all previous steps are passed to the agent at future iterations, but in long-running cases that could lead to an overly large amount of steps that you may want to trim\n",
"\n",
"This examples builds off the base agent executor. It is highly recommended you learn about that executor before going through this notebook. You can find documentation for that example [here](./base.ipynb).\n",
"\n",
"Any modifications of that example are called below with **MODIFICATION**, so if you are looking for the differences you can just search for that."
"os.environ[\"LANGCHAIN_API_KEY\"] = getpass.getpass(\"LangSmith API Key:\")"
]
},
{
"cell_type": "markdown",
"id": "5dace4a9-7c9e-4da2-bf7b-e58d0d05581e",
"metadata": {},
"source": [
"## Create the LangChain agent\n",
"\n",
"First, we will create the LangChain agent. For more information on LangChain agents, see [this documentation](https://python.langchain.com/v0.2/docs/concepts/#agents)"
"We now define the graph state. The state for the traditional LangChain agent has a few attributes:\n",
"\n",
"1. `input`: This is the input string representing the main ask from the user, passed in as input.\n",
"2. `chat_history`: This is any previous conversation messages, also passed in as input.\n",
"3. `intermediate_steps`: This is list of actions and corresponding observations that the agent takes over time. This is updated each iteration of the agent.\n",
"4. `agent_outcome`: This is the response from the agent, either an AgentAction or AgentFinish. The AgentExecutor should finish when this is an AgentFinish, otherwise it should call the requested tools.\n"
"We now need to define a few different nodes in our graph.\n",
"In `langgraph`, a node can be either a function or a [runnable](https://python.langchain.com/v0.2/docs/concepts/#langchain-expression-language-lcel).\n",
"There are two main nodes we need for this:\n",
"\n",
"1. The agent: responsible for deciding what (if any) actions to take.\n",
"2. A function to invoke tools: if the agent decides to take an action, this node will then execute that action.\n",
"\n",
"We will also need to define some edges.\n",
"Some of these edges may be conditional.\n",
"The reason they are conditional is that based on the output of a node, one of several paths may be taken.\n",
"The path that is taken is not known until that node is run (the LLM decides).\n",
"\n",
"1. Conditional Edge: after the agent is called, we should either:\n",
" a. If the agent said to take an action, then the function to invoke tools should be called\n",
" b. If the agent said that it was finished, then it should finish\n",
"2. Normal Edge: after the tools are invoked, it should always go back to the agent to decide what to do next\n",
"\n",
"Let's define the nodes, as well as a function to decide how what conditional edge to take."
"# This a helper class we have that is useful for running tools\n",
"# It takes in an agent action and calls that tool and returns the result\n",
"tool_executor = ToolExecutor(tools)"
]
},
{
"cell_type": "markdown",
"id": "4c804a34-d384-4ca9-b9fc-dc86d678ab39",
"metadata": {},
"source": [
"**MODIFICATION**\n",
"\n",
"Here, we modify the agent to only look at the last five intermediate steps. This is a relatively simple example of shortening the intermediate step history."
"# Define logic that will be used to determine which conditional edge to go down\n",
"def should_continue(data):\n",
" # If the agent outcome is an AgentFinish, then we return `exit` string\n",
" # This will be used when setting up the graph to define the flow\n",
" if isinstance(data[\"agent_outcome\"], AgentFinish):\n",
" return \"end\"\n",
" # Otherwise, an AgentAction is returned\n",
" # Here we return `continue` string\n",
" # This will be used when setting up the graph to define the flow\n",
" else:\n",
" return \"continue\""
]
},
{
"cell_type": "markdown",
"id": "c0b211f4-0c5c-4792-b18d-cd70907c71e7",
"metadata": {},
"source": [
"## Define the graph\n",
"\n",
"We can now put it all together and define the graph!"
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "c4054dde-4618-49b7-998a-daa0c1d6d6c0",
"metadata": {},
"outputs": [],
"source": [
"from langgraph.graph import END, StateGraph\n",
"\n",
"# Define a new graph\n",
"workflow = StateGraph(AgentState)\n",
"\n",
"# Define the two nodes we will cycle between\n",
"workflow.add_node(\"agent\", run_agent)\n",
"workflow.add_node(\"action\", execute_tools)\n",
"\n",
"# Set the entrypoint as `agent`\n",
"# This means that this node is the first one called\n",
"workflow.set_entry_point(\"agent\")\n",
"\n",
"# We now add a conditional edge\n",
"workflow.add_conditional_edges(\n",
" # First, we define the start node. We use `agent`.\n",
" # This means these are the edges taken after the `agent` node is called.\n",
" \"agent\",\n",
" # Next, we pass in the function that will determine which node is called next.\n",
" should_continue,\n",
" # Finally we pass in a mapping.\n",
" # The keys are strings, and the values are other nodes.\n",
" # END is a special node marking that the graph should finish.\n",
" # What will happen is we will call `should_continue`, and then the output of that\n",
" # will be matched against the keys in this mapping.\n",
" # Based on which one it matches, that node will then be called.\n",
" {\n",
" # If `tools`, then we call the tool node.\n",
" \"continue\": \"action\",\n",
" # Otherwise we finish.\n",
" \"end\": END,\n",
" },\n",
")\n",
"\n",
"# We now add a normal edge from `tools` to `agent`.\n",
"# This means that after `tools` is called, `agent` node is called next.\n",
"workflow.add_edge(\"action\", \"agent\")\n",
"\n",
"# Finally, we compile it!\n",
"# This compiles it into a LangChain Runnable,\n",
"# meaning you can use it as you would any other runnable\n",
"app = workflow.compile()"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "214ae46e-c297-465d-86db-2b0312ed3530",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'agent_outcome': AgentActionMessageLog(tool='tavily_search_results_json', tool_input={'query': 'weather in San Francisco'}, log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\", message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})])}\n",
"----\n",
"{'intermediate_steps': [(AgentActionMessageLog(tool='tavily_search_results_json', tool_input={'query': 'weather in San Francisco'}, log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\", message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})]), \"[{'url': 'https://en.climate-data.org/north-america/united-states-of-america/california/san-francisco-385/t/january-1/', 'content': 'San Francisco Weather in January San Francisco weather in January San Francisco weather by month // weather averages 9.6 (49.2) 6.2 (43.2) 14 (57.3) 113 San Francisco weather in January // weather averages Airport close to San Francisco you can find all information about the weather in San Francisco in January:Data: 1991 - 2021 Min. Temperature °C (°F), Max. Temperature °C (°F), Precipitation / Rainfall mm (in), Humidity, Rainy days. Data: 1999 - 2019: avg. Sun hours San Francisco weather and climate for further months San Francisco in February San Francisco in March San Francisco in April San Francisco in May San Francisco in June San Francisco in July'}]\")]}\n",
"----\n",
"{'agent_outcome': AgentFinish(return_values={'output': \"The weather in San Francisco varies by month. In January, the average minimum temperature is 9.6°C (49.2°F), and the average maximum temperature is 14°C (57.3°F). The city experiences an average of 113mm of precipitation and has an average of 6 rainy days in January. If you'd like to know more about the weather in other months, feel free to ask!\"}, log=\"The weather in San Francisco varies by month. In January, the average minimum temperature is 9.6°C (49.2°F), and the average maximum temperature is 14°C (57.3°F). The city experiences an average of 113mm of precipitation and has an average of 6 rainy days in January. If you'd like to know more about the weather in other months, feel free to ask!\")}\n",
"----\n",
"{'input': 'what is the weather in sf', 'chat_history': [], 'agent_outcome': AgentFinish(return_values={'output': \"The weather in San Francisco varies by month. In January, the average minimum temperature is 9.6°C (49.2°F), and the average maximum temperature is 14°C (57.3°F). The city experiences an average of 113mm of precipitation and has an average of 6 rainy days in January. If you'd like to know more about the weather in other months, feel free to ask!\"}, log=\"The weather in San Francisco varies by month. In January, the average minimum temperature is 9.6°C (49.2°F), and the average maximum temperature is 14°C (57.3°F). The city experiences an average of 113mm of precipitation and has an average of 6 rainy days in January. If you'd like to know more about the weather in other months, feel free to ask!\"), 'intermediate_steps': [(AgentActionMessageLog(tool='tavily_search_results_json', tool_input={'query': 'weather in San Francisco'}, log=\"\\nInvoking: `tavily_search_results_json` with `{'query': 'weather in San Francisco'}`\\n\\n\\n\", message_log=[AIMessage(content='', additional_kwargs={'function_call': {'arguments': '{\"query\":\"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}})]), \"[{'url': 'https://en.climate-data.org/north-america/united-states-of-america/california/san-francisco-385/t/january-1/', 'content': 'San Francisco Weather in January San Francisco weather in January San Francisco weather by month // weather averages 9.6 (49.2) 6.2 (43.2) 14 (57.3) 113 San Francisco weather in January // weather averages Airport close to San Francisco you can find all information about the weather in San Francisco in January:Data: 1991 - 2021 Min. Temperature °C (°F), Max. Temperature °C (°F), Precipitation / Rainfall mm (in), Humidity, Rainy days. Data: 1999 - 2019: avg. Sun hours San Francisco weather and climate for further months San Francisco in February San Francisco in March San Francisco in April San Francisco in May San Francisco in June San Francisco in July'}]\")]}\n",
"----\n"
]
}
],
"source": [
"inputs = {\"input\": \"what is the weather in sf\", \"chat_history\": []}\n",
"os.environ[\"LANGCHAIN_API_KEY\"] = getpass.getpass(\"LangSmith API Key:\")"
]
},
{
"cell_type": "markdown",
"id": "21ac643b-cb06-4724-a80c-2862ba4773f1",
"metadata": {},
"source": [
"## Set up the tools\n",
"\n",
"We will first define the tools we want to use.\n",
"For this simple example, we will use create a placeholder search engine.\n",
"However, it is really easy to create your own tools - see documentation [here](https://python.langchain.com/v0.2/docs/how_to/custom_tools) on how to do that.\n",
"\n",
"**MODIFICATION**\n",
"\n",
"We don't need a ToolExecutor when using ToolNode.\n"
"After we've done this, we should make sure the model knows that it has these tools available to call.\n",
"We can do this by converting the LangChain tools into the format for OpenAI function calling, and then bind them to the model class.\n"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "cd3cbae5-d92c-4559-a4aa-44721b80d107",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"/Users/nuno/dev/langgraph/.venv/lib/python3.11/site-packages/langchain_core/_api/beta_decorator.py:87: LangChainBetaWarning: The method `ChatAnthropic.bind_tools` is in beta. It is actively being worked on, so the API may change.\n",
"We now need to define a few different nodes in our graph.\n",
"In `langgraph`, a node can be either a function or a [runnable](https://python.langchain.com/v0.2/docs/concepts/#langchain-expression-language-lcel).\n",
"There are two main nodes we need for this:\n",
"\n",
"1. The agent: responsible for deciding what (if any) actions to take.\n",
"2. **MODIFICATION** The prebuilt ToolNode, given the list of tools. This will take tool calls from the most recent AIMessage, execute them, and return the result as ToolMessages.\n",
"\n",
"We will also need to define some edges.\n",
"Some of these edges may be conditional.\n",
"The reason they are conditional is that based on the output of a node, one of several paths may be taken.\n",
"The path that is taken is not known until that node is run (the LLM decides).\n",
"\n",
"1. Conditional Edge: after the agent is called, we should either:\n",
" a. If the agent said to take an action, then the function to invoke tools should be called\n",
" b. If the agent said that it was finished, then it should finish\n",
"2. Normal Edge: after the tools are invoked, it should always go back to the agent to decide what to do next\n",
"\n",
"Let's define the nodes, as well as a function to decide how what conditional edge to take.\n"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "3b541bb9-900c-40d0-964d-7b5dfee30667",
"metadata": {},
"outputs": [],
"source": [
"from langgraph.prebuilt import ToolNode\n",
"\n",
"\n",
"# Define the function that determines whether to continue or not\n",
"def should_continue(state):\n",
" messages = state[\"messages\"]\n",
" last_message = messages[-1]\n",
" # If there are no tool calls, then we finish\n",
" if not last_message.tool_calls:\n",
" return \"end\"\n",
" # Otherwise if there is, we continue\n",
" else:\n",
" return \"continue\"\n",
"\n",
"\n",
"# Define the function that calls the model\n",
"def call_model(state):\n",
" messages = state[\"messages\"]\n",
" response = model.invoke(messages)\n",
" # We return a list, because this will get added to the existing list\n",
" return {\"messages\": [response]}\n",
"\n",
"\n",
"# Define the function to execute tools\n",
"tool_node = ToolNode(tools)"
]
},
{
"cell_type": "markdown",
"id": "ffd6e892-946c-4899-8cc0-7c9291c1f73b",
"metadata": {},
"source": [
"## Define the graph\n",
"\n",
"We can now put it all together and define the graph!"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "813ae66c-3b58-4283-a02a-36da72a2ab90",
"metadata": {},
"outputs": [],
"source": [
"from langgraph.graph import END, StateGraph\n",
"\n",
"# Define a new graph\n",
"workflow = StateGraph(AgentState)\n",
"\n",
"# Define the two nodes we will cycle between\n",
"workflow.add_node(\"agent\", call_model)\n",
"workflow.add_node(\"action\", tool_node)\n",
"\n",
"# Set the entrypoint as `agent`\n",
"# This means that this node is the first one called\n",
"workflow.set_entry_point(\"agent\")\n",
"\n",
"# We now add a conditional edge\n",
"workflow.add_conditional_edges(\n",
" # First, we define the start node. We use `agent`.\n",
" # This means these are the edges taken after the `agent` node is called.\n",
" \"agent\",\n",
" # Next, we pass in the function that will determine which node is called next.\n",
" should_continue,\n",
" # Finally we pass in a mapping.\n",
" # The keys are strings, and the values are other nodes.\n",
" # END is a special node marking that the graph should finish.\n",
" # What will happen is we will call `should_continue`, and then the output of that\n",
" # will be matched against the keys in this mapping.\n",
" # Based on which one it matches, that node will then be called.\n",
" {\n",
" # If `tools`, then we call the tool node.\n",
" \"continue\": \"action\",\n",
" # Otherwise we finish.\n",
" \"end\": END,\n",
" },\n",
")\n",
"\n",
"# We now add a normal edge from `tools` to `agent`.\n",
"# This means that after `tools` is called, `agent` node is called next.\n",
"workflow.add_edge(\"action\", \"agent\")\n",
"\n",
"# Finally, we compile it!\n",
"# This compiles it into a LangChain Runnable,\n",
"# meaning you can use it as you would any other runnable\n",
"app = workflow.compile()"
]
},
{
"cell_type": "markdown",
"id": "547c3931-3dae-4281-ad4e-4b51305594d4",
"metadata": {},
"source": [
"## Use it!\n",
"\n",
"We can now use it!\n",
"This now exposes the [same interface](https://python.langchain.com/v0.2/docs/concepts/#langchain-expression-language-lcel) as all other LangChain runnables."
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "8edb04b9-40b6-46f1-a7a8-4b2d8aba7752",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'messages': [HumanMessage(content='what is the weather in sf'),\n",
" AIMessage(content=[{'text': '<thinking>\\nThe relevant tool to answer this question is tavily_search_results_json, which can provide comprehensive information about current events like weather.\\n\\nTo call this function, I need to provide a value for the required \"query\" parameter. The user\\'s request directly specifies they want to know the weather in \"sf\", which I can reasonably infer refers to San Francisco.\\n\\nTherefore, I have enough information to populate the required parameter:\\nquery = \"weather in San Francisco\"\\n\\n</thinking>', 'type': 'text'}, {'id': 'toolu_0183a3MorRJu43zykiCWKAyo', 'input': {'query': 'weather in San Francisco'}, 'name': 'tavily_search_results_json', 'type': 'tool_use'}], response_metadata={'id': 'msg_01Lg8ZNFNwbDXz9VfxZyRCSb', 'model': 'claude-3-opus-20240229', 'stop_reason': 'tool_use', 'stop_sequence': None, 'usage': {'input_tokens': 507, 'output_tokens': 166}}, id='run-587209cf-1406-47f1-9476-73f9c75f4650-0', tool_calls=[{'name': 'tavily_search_results_json', 'args': {'query': 'weather in San Francisco'}, 'id': 'toolu_0183a3MorRJu43zykiCWKAyo'}]),\n",
" AIMessage(content=\"<search_quality_reflection>\\nThe search results provide a comprehensive and up-to-date weather report for San Francisco, including key details like the current temperature, weather conditions, wind, humidity, and more. This should be sufficient to fully answer the question of what the current weather is like in San Francisco.\\n</search_quality_reflection>\\n\\n<search_quality_score>5</search_quality_score>\\n\\n<result>\\nAccording to the current weather report, the weather in San Francisco right now is:\\n\\nTemperature: 63°F (17.2°C)\\nConditions: Partly cloudy \\nWind: 34.9 mph (56.2 km/h) winds from the west\\nHumidity: 60%\\n\\nIt feels like 63°F (17.2°C). Visibility is good at 9 miles (16 km). The UV index is moderate at 4.0 out of 11. \\n\\nOverall, it's a mild spring day in San Francisco with some cloud cover and breezy conditions. A light jacket or sweater should suffice for being outdoors.\\n</result>\", response_metadata={'id': 'msg_01LS72RMeicMF1xT7enopKpJ', 'model': 'claude-3-opus-20240229', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'input_tokens': 1097, 'output_tokens': 251}}, id='run-794deb88-bea5-4d0d-93db-bf5dc38445f0-0')]}"
"inputs = {\"messages\": [HumanMessage(content=\"what is the weather in sf\")]}\n",
"app.invoke(inputs)"
]
},
{
"cell_type": "markdown",
"id": "5a9e8155-70c5-4973-912c-dc55104b2acf",
"metadata": {},
"source": [
"This may take a little bit - it's making a few calls behind the scenes.\n",
"In order to start seeing some intermediate results as they happen, we can use streaming - see below for more information on that.\n",
"\n",
"## Streaming\n",
"\n",
"LangGraph has support for several different types of streaming.\n",
"\n",
"### Streaming Node Output\n",
"\n",
"One of the benefits of using LangGraph is that it is easy to stream output as it's produced by each node.\n"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "f544977e-31f7-41f0-88c4-ec9c27b8cecb",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Output from node 'agent':\n",
"---\n",
"{'messages': [AIMessage(content=[{'text': '<thinking>\\nThe relevant tool to answer this question is tavily_search_results_json, which can provide comprehensive results about current events like weather.\\n\\nTo call this function, I need to provide a value for the required \"query\" parameter. The user\\'s request directly specifies the query to search for: \"weather in sf\". \"sf\" here likely refers to San Francisco.\\n\\nSince I have a value for the required parameter, I can proceed with the function call.\\n</thinking>', 'type': 'text'}, {'id': 'toolu_01XgUtdMt17UaBS8BUN2ZRyn', 'input': {'query': 'weather in San Francisco'}, 'name': 'tavily_search_results_json', 'type': 'tool_use'}], response_metadata={'id': 'msg_01SyKFjD9dxUNxwTQ5FiT3Yr', 'model': 'claude-3-opus-20240229', 'stop_reason': 'tool_use', 'stop_sequence': None, 'usage': {'input_tokens': 507, 'output_tokens': 162}}, id='run-42b25509-f322-4c4b-9817-f9ae154b8293-0', tool_calls=[{'name': 'tavily_search_results_json', 'args': {'query': 'weather in San Francisco'}, 'id': 'toolu_01XgUtdMt17UaBS8BUN2ZRyn'}])]}\n",
"{'messages': [AIMessage(content='<search_quality_reflection>\\nThe search results provide a comprehensive and up-to-date weather report for San Francisco, including key details like temperature, conditions, wind, humidity, and more. This should be sufficient to fully answer the question of what the current weather is like in San Francisco.\\n</search_quality_reflection>\\n<search_quality_score>5</search_quality_score>\\n\\n<result>\\nAccording to the latest weather report, the current weather in San Francisco is:\\n\\nTemperature: 60.1°F (15.6°C)\\nConditions: Partly cloudy \\nWind: 4.3 mph (6.8 km/h) from the NE\\nHumidity: 78%\\nPrecipitation: 0 inches\\nVisibility: 9 miles\\nUV Index: 5.0\\n\\nIt feels like 60.1°F (15.6°C). The report indicates it is a partly cloudy day with no rain expected. Winds are light out of the northeast.\\n</result>', response_metadata={'id': 'msg_01X8S82ECeXU8px2TpMPfkce', 'model': 'claude-3-opus-20240229', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'input_tokens': 1094, 'output_tokens': 232}}, id='run-772e7225-dc58-4b63-a0d7-6d7d39e3b059-0')]}\n",
"\n",
"---\n",
"\n"
]
}
],
"source": [
"inputs = {\"messages\": [HumanMessage(content=\"what is the weather in sf\")]}\n",
"for output in app.stream(inputs):\n",
" # stream() yields dictionaries with output keyed by node name\n",
"We can now use the high level interface to create the executor"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "32b4ae66-f667-4a8b-a602-503fd0effcd9",
"metadata": {},
"outputs": [],
"source": [
"app = create_react_agent(model, tools=tools)"
]
},
{
"cell_type": "markdown",
"id": "d63dbfc7-a5c1-4a03-991c-f0789ba52c52",
"metadata": {},
"source": [
"We can now invoke this executor. The input to this must be a dictionary with a single `messages` key that contains a list of messages."
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "0abc5655-d772-450c-832f-1fee1111a5f6",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_eI2B853W8Jrm8IvmwEafikFv', 'function': {'arguments': '{\"query\": \"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}, 'type': 'function'}, {'id': 'call_Aky1m2Z5dvUcHKyha7r5s3Wj', 'function': {'arguments': '{\"query\": \"weather in Los Angeles\"}', 'name': 'tavily_search_results_json'}, 'type': 'function'}]})]}\n",
"----\n",
"{'messages': [ToolMessage(content=\"[{'url': 'https://www.wunderground.com/forecast/us/ca/san-francisco', 'content': 'Get the latest weather information for San Francisco, CA, including temperature, precipitation, wind speed, and humidity. See the hourly and 10-day forecast for the South of Market station and other nearby weather stations.'}]\", tool_call_id='call_eI2B853W8Jrm8IvmwEafikFv'), ToolMessage(content=\"[{'url': 'https://www.accuweather.com/en/us/los-angeles/90012/hourly-weather-forecast/347625', 'content': 'Get the latest hourly weather updates for Los Angeles, CA, including rain alerts, air quality, wind speed and direction, humidity, and cloud cover. See the forecast for the next eight hours and plan your activities accordingly.'}]\", tool_call_id='call_Aky1m2Z5dvUcHKyha7r5s3Wj')]}\n",
"----\n",
"{'messages': [AIMessage(content='The weather in San Francisco can be found [here](https://www.wunderground.com/forecast/us/ca/san-francisco), which includes information on temperature, precipitation, wind speed, and humidity.\\n\\nFor Los Angeles, you can check the hourly weather updates [here](https://www.accuweather.com/en/us/los-angeles/90012/hourly-weather-forecast/347625), which includes details on rain alerts, air quality, wind speed and direction, humidity, and cloud cover.')]}\n",
"----\n",
"{'messages': [HumanMessage(content='what is the weather in sf and la'), AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_eI2B853W8Jrm8IvmwEafikFv', 'function': {'arguments': '{\"query\": \"weather in San Francisco\"}', 'name': 'tavily_search_results_json'}, 'type': 'function'}, {'id': 'call_Aky1m2Z5dvUcHKyha7r5s3Wj', 'function': {'arguments': '{\"query\": \"weather in Los Angeles\"}', 'name': 'tavily_search_results_json'}, 'type': 'function'}]}), ToolMessage(content=\"[{'url': 'https://www.wunderground.com/forecast/us/ca/san-francisco', 'content': 'Get the latest weather information for San Francisco, CA, including temperature, precipitation, wind speed, and humidity. See the hourly and 10-day forecast for the South of Market station and other nearby weather stations.'}]\", tool_call_id='call_eI2B853W8Jrm8IvmwEafikFv'), ToolMessage(content=\"[{'url': 'https://www.accuweather.com/en/us/los-angeles/90012/hourly-weather-forecast/347625', 'content': 'Get the latest hourly weather updates for Los Angeles, CA, including rain alerts, air quality, wind speed and direction, humidity, and cloud cover. See the forecast for the next eight hours and plan your activities accordingly.'}]\", tool_call_id='call_Aky1m2Z5dvUcHKyha7r5s3Wj'), AIMessage(content='The weather in San Francisco can be found [here](https://www.wunderground.com/forecast/us/ca/san-francisco), which includes information on temperature, precipitation, wind speed, and humidity.\\n\\nFor Los Angeles, you can check the hourly weather updates [here](https://www.accuweather.com/en/us/los-angeles/90012/hourly-weather-forecast/347625), which includes details on rain alerts, air quality, wind speed and direction, humidity, and cloud cover.')]}\n",
"----\n"
]
}
],
"source": [
"inputs = {\"messages\": [HumanMessage(content=\"what is the weather in sf and la\")]}\n",
"# (Deprecated) Chat Executor: with function calling\n",
"\n",
"The function calling executor is deprecated in favor of [create_react_agent](../chat_agent_executor_with_function_calling/high-level-tools.ipynb).\n",
"This was done to better align with the underlying model providers' migration from \"function calling\" to \"tool calling\", which typically supports parallel tool usage."
"os.environ[\"LANGCHAIN_API_KEY\"] = getpass.getpass(\"LangSmith API Key:\")"
]
},
{
"cell_type": "markdown",
"id": "21ac643b-cb06-4724-a80c-2862ba4773f1",
"metadata": {},
"source": [
"## Set up the tools\n",
"\n",
"We will first define the tools we want to use.\n",
"For this simple example, we will use a built-in search tool via Tavily.\n",
"However, it is really easy to create your own tools - see documentation [here](https://python.langchain.com/v0.2/docs/how_to/custom_tools) on how to do that.\n",
"\n",
"**MODIFICATION**\n",
"\n",
"We don't need a ToolExecutor when using ToolNode.\n"
"We now need to define a few different nodes in our graph.\n",
"In `langgraph`, a node can be either a function or a [runnable](https://python.langchain.com/v0.2/docs/concepts/#langchain-expression-language-lcel).\n",
"There are two main nodes we need for this:\n",
"\n",
"1. The agent: responsible for deciding what (if any) actions to take.\n",
"2. **MODIFICATION** The prebuilt ToolNode, given the list of tools. This will take tool calls from the most recent AIMessage, execute them, and return the result as ToolMessages.\n",
"\n",
"We will also need to define some edges.\n",
"Some of these edges may be conditional.\n",
"The reason they are conditional is that based on the output of a node, one of several paths may be taken.\n",
"The path that is taken is not known until that node is run (the LLM decides).\n",
"\n",
"1. Conditional Edge: after the agent is called, we should either:\n",
" a. If the agent said to take an action, then the function to invoke tools should be called\n",
" b. If the agent said that it was finished, then it should finish\n",
"2. Normal Edge: after the tools are invoked, it should always go back to the agent to decide what to do next\n",
"\n",
"Let's define the nodes, as well as a function to decide how what conditional edge to take.\n"
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "3b541bb9-900c-40d0-964d-7b5dfee30667",
"metadata": {},
"outputs": [],
"source": [
"from langgraph.prebuilt import ToolNode\n",
"\n",
"\n",
"# Define the function that determines whether to continue or not\n",
"def should_continue(state):\n",
" messages = state[\"messages\"]\n",
" last_message = messages[-1]\n",
" # If there are no tool calls, then we finish\n",
" if not last_message.tool_calls:\n",
" return \"end\"\n",
" # Otherwise if there is, we continue\n",
" else:\n",
" return \"continue\"\n",
"\n",
"\n",
"# Define the function that calls the model\n",
"def call_model(state):\n",
" messages = state[\"messages\"]\n",
" response = model.invoke(messages)\n",
" # We return a list, because this will get added to the existing list\n",
" return {\"messages\": [response]}\n",
"\n",
"\n",
"# Define the function to execute tools\n",
"tool_node = ToolNode(tools)"
]
},
{
"cell_type": "markdown",
"id": "ffd6e892-946c-4899-8cc0-7c9291c1f73b",
"metadata": {},
"source": [
"## Define the graph\n",
"\n",
"We can now put it all together and define the graph!"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "813ae66c-3b58-4283-a02a-36da72a2ab90",
"metadata": {},
"outputs": [],
"source": [
"from langgraph.graph import END, StateGraph\n",
"\n",
"# Define a new graph\n",
"workflow = StateGraph(AgentState)\n",
"\n",
"# Define the two nodes we will cycle between\n",
"workflow.add_node(\"agent\", call_model)\n",
"workflow.add_node(\"action\", tool_node)\n",
"\n",
"# Set the entrypoint as `agent`\n",
"# This means that this node is the first one called\n",
"workflow.set_entry_point(\"agent\")\n",
"\n",
"# We now add a conditional edge\n",
"workflow.add_conditional_edges(\n",
" # First, we define the start node. We use `agent`.\n",
" # This means these are the edges taken after the `agent` node is called.\n",
" \"agent\",\n",
" # Next, we pass in the function that will determine which node is called next.\n",
" should_continue,\n",
" # Finally we pass in a mapping.\n",
" # The keys are strings, and the values are other nodes.\n",
" # END is a special node marking that the graph should finish.\n",
" # What will happen is we will call `should_continue`, and then the output of that\n",
" # will be matched against the keys in this mapping.\n",
" # Based on which one it matches, that node will then be called.\n",
" {\n",
" # If `tools`, then we call the tool node.\n",
" \"continue\": \"action\",\n",
" # Otherwise we finish.\n",
" \"end\": END,\n",
" },\n",
")\n",
"\n",
"# We now add a normal edge from `tools` to `agent`.\n",
"# This means that after `tools` is called, `agent` node is called next.\n",
"workflow.add_edge(\"action\", \"agent\")\n",
"\n",
"# Finally, we compile it!\n",
"# This compiles it into a LangChain Runnable,\n",
"# meaning you can use it as you would any other runnable\n",
"app = workflow.compile()"
]
},
{
"cell_type": "markdown",
"id": "547c3931-3dae-4281-ad4e-4b51305594d4",
"metadata": {},
"source": [
"## Use it!\n",
"\n",
"We can now use it!\n",
"This now exposes the [same interface](https://python.langchain.com/v0.2/docs/concepts/#langchain-expression-language-lcel) as all other LangChain runnables."
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "8edb04b9-40b6-46f1-a7a8-4b2d8aba7752",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'messages': [HumanMessage(content='what is the weather in sf'),\n",
" AIMessage(content='The current weather in San Francisco is as follows:\\n- Temperature: 15.0°C (59.0°F)\\n- Condition: Partly cloudy\\n- Wind: 3.8 mph from the North\\n- Humidity: 78%\\n- Visibility: 16.0 km (9.0 miles)\\n- UV Index: 4.0\\n\\nFor more details, you can visit [Weather API](https://www.weatherapi.com/).', response_metadata={'token_usage': {'completion_tokens': 93, 'prompt_tokens': 465, 'total_tokens': 558}, 'model_name': 'gpt-3.5-turbo', 'system_fingerprint': 'fp_b28b39ffa8', 'finish_reason': 'stop', 'logprobs': None}, id='run-923bcbd2-3c79-4696-8f9e-5142b50b20cf-0')]}"
"{'messages': [AIMessage(content='The current weather in San Francisco is partly cloudy with a temperature of 59°F (15°C). The wind speed is 6.1 km/h coming from the north. The humidity is at 78%, and the visibility is 16.0 km.', response_metadata={'token_usage': {'completion_tokens': 53, 'prompt_tokens': 465, 'total_tokens': 518}, 'model_name': 'gpt-3.5-turbo', 'system_fingerprint': 'fp_b28b39ffa8', 'finish_reason': 'stop', 'logprobs': None}, id='run-8875456d-e31e-42b0-b2af-bdc1a9cfccfe-0')]}\n",
"\n",
"---\n",
"\n"
]
}
],
"source": [
"inputs = {\"messages\": [HumanMessage(content=\"what is the weather in sf\")]}\n",
"for output in app.stream(inputs):\n",
" # stream() yields dictionaries with output keyed by node name\n",
" for key, value in output.items():\n",
" print(f\"Output from node '{key}':\")\n",
" print(\"---\")\n",
" print(value)\n",
" print(\"\\n---\\n\")"
]
},
{
"cell_type": "markdown",
"id": "2a1b56c5-bd61-4192-8bdb-458a1e9f0159",
"metadata": {},
"source": [
"### Streaming LLM Tokens\n",
"\n",
"You can also access the LLM tokens as they are produced by each node. \n",
"In this case only the \"agent\" node produces LLM tokens.\n",
"In order for this to work properly, you must be using an LLM that supports streaming as well as have set it when constructing the LLM (e.g. `ChatOpenAI(model=\"gpt-3.5-turbo-1106\", streaming=True)`)\n"
"# Chat Bot Evaluation as Multi-agent Simulation\n",
"\n",
"When building a chat bot, such as a customer support assistant, it can be hard to properly evaluate your bot's performance. It's time-consuming to have to manually interact with it intensively for each code change.\n",
"\n",
"One way to make the evaluation process easier and more reproducible is to simulate a user interaction.\n",
"\n",
"With LangGraph, it's easy to set this up. Below is an example of how to create a \"virtual user\" to simulate a conversation.\n",
"\n",
"The overall simulation looks something like this:\n",
"Next, we will define our chat bot. For this notebook, we assume the bot's API accepts a list of messages and responds with a message. If you want to update this, all you'll have to change is this section and the \"get_messages_for_agent\" function in \n",
"the simulator below.\n",
"\n",
"The implementation within `my_chat_bot` is configurable and can even be run on another system (e.g., if your system isn't running in python)."
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "828479af-cf9c-4888-a365-599643a96b55",
"metadata": {},
"outputs": [],
"source": [
"from typing import List\n",
"\n",
"import openai\n",
"\n",
"\n",
"# This is flexible, but you can define your agent here, or call your agent API here.\n",
"AIMessage(content='Hi, I would like to request a refund for a trip I took with your airline company to Alaska. Is it possible to get a refund for that trip?')"
"messages = [HumanMessage(content=\"Hi! How can I help you?\")]\n",
"simulated_user.invoke({\"messages\": messages})"
]
},
{
"cell_type": "markdown",
"id": "321312b4-a1f0-4454-a481-fdac4e37cb7d",
"metadata": {},
"source": [
"## 3. Define the Agent Simulation\n",
"\n",
"The code below creates a LangGraph workflow to run the simulation. The main components are:\n",
"\n",
"1. The two nodes: one for the simulated user, the other for the chat bot.\n",
"2. The graph itself, with a conditional stopping criterion.\n",
"\n",
"Read the comments in the code below for more information.\n"
]
},
{
"cell_type": "markdown",
"id": "65bc4446-462b-4ee8-b017-2862fbbdfaf5",
"metadata": {},
"source": [
"**Nodes**\n",
"\n",
"First, we define the nodes in the graph. These should take in a list of messages and return a list of messages to ADD to the state.\n",
"These will be thing wrappers around the chat bot and simulated user we have above.\n",
"\n",
"**Note:** one tricky thing here is which messages are which. Because both the chat bot AND our simulated user are both LLMs, both of them will resond with AI messages. Our state will be a list of alternating Human and AI messages. This means that for one of the nodes, there will need to be some logic that flips the AI and human roles. In this example, we will assume that HumanMessages are messages from the simulated user. This means that we need some logic in the simulated user node to swap AI and Human messages.\n",
" # This response is an AI message - we need to flip this to be a human message\n",
" return HumanMessage(content=response.content)"
]
},
{
"cell_type": "markdown",
"id": "a48d8a3e-9171-4c43-a595-44d312722148",
"metadata": {},
"source": [
"**Edges**\n",
"\n",
"We now need to define the logic for the edges. The main logic occurs after the simulated user goes, and it should lead to one of two outcomes:\n",
"\n",
"- Either we continue and call the customer support bot\n",
"- Or we finish and the conversation is over\n",
"\n",
"So what is the logic for the conversation being over? We will define that as either the Human chatbot responds with `FINISHED` (see the system prompt) OR the conversation is more than 6 messages long (this is an arbitrary number just to keep this example short)."
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "28004fbf-a2f3-46b7-bde7-46c7adaf97fb",
"metadata": {},
"outputs": [],
"source": [
"def should_continue(messages):\n",
" if len(messages) > 6:\n",
" return \"end\"\n",
" elif messages[-1].content == \"FINISHED\":\n",
" return \"end\"\n",
" else:\n",
" return \"continue\""
]
},
{
"cell_type": "markdown",
"id": "d0856d4f-9334-4f28-944b-06d303e913a4",
"metadata": {},
"source": [
"**Graph**\n",
"\n",
"We can now define the graph that sets up the simulation!"
" # If the finish criteria are met, we will stop the simulation,\n",
" # otherwise, the virtual user's message will be sent to your chat bot\n",
" {\n",
" \"end\": END,\n",
" \"continue\": \"chat_bot\",\n",
" },\n",
")\n",
"# The input will first go to your chat bot\n",
"graph_builder.set_entry_point(\"chat_bot\")\n",
"simulation = graph_builder.compile()"
]
},
{
"cell_type": "markdown",
"id": "2e0bd26e-8c1d-471d-9fef-d95dc0163491",
"metadata": {},
"source": [
"## 4. Run Simulation\n",
"\n",
"Now we can evaluate our chat bot! We can invoke it with empty messages (this will simulate letting the chat bot start the initial conversation)"
]
},
{
"cell_type": "code",
"execution_count": 11,
"id": "32848c2e-be82-46f3-81db-b23fea45461c",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'chat_bot': AIMessage(content='How may I assist you today regarding your flight or any other concerns?')}\n",
"----\n",
"{'user': HumanMessage(content='Hi, my name is Harrison. I am reaching out to request a refund for a trip I took to Alaska with your airline company. The trip occurred about 5 years ago. I would like to receive a refund for the entire amount I paid for the trip. Can you please assist me with this?')}\n",
"----\n",
"{'chat_bot': AIMessage(content=\"Hello, Harrison. Thank you for reaching out to us. I understand you would like to request a refund for a trip you took to Alaska five years ago. I'm afraid that our refund policy typically has a specific timeframe within which refund requests must be made. Generally, refund requests need to be submitted within 24 to 48 hours after the booking is made, or in certain cases, within a specified cancellation period.\\n\\nHowever, I will do my best to assist you. Could you please provide me with some additional information? Can you recall any specific details about the booking, such as the flight dates, booking reference or confirmation number? This will help me further look into the possibility of processing a refund for you.\")}\n",
"----\n",
"{'user': HumanMessage(content=\"Hello, thank you for your response. I apologize for not requesting the refund earlier. Unfortunately, I don't have the specific details such as the flight dates, booking reference, or confirmation number at the moment. Is there any other way we can proceed with the refund request without these specific details? I would greatly appreciate your assistance in finding a solution.\")}\n",
"----\n",
"{'chat_bot': AIMessage(content=\"I understand the situation, Harrison. Without specific details like flight dates, booking reference, or confirmation number, it becomes challenging to locate and process the refund accurately. However, I can still try to help you.\\n\\nTo proceed further, could you please provide me with any additional information you might remember? This could include the approximate date of travel, the departure and arrival airports, the names of the passengers, or any other relevant details related to the booking. The more information you can provide, the better we can investigate the possibility of processing a refund for you.\\n\\nAdditionally, do you happen to have any documentation related to your trip, such as receipts, boarding passes, or emails from our airline? These documents could assist in verifying your trip and processing the refund request.\\n\\nI apologize for any inconvenience caused, and I'll do my best to assist you further based on the information you can provide.\")}\n",
"----\n",
"{'user': HumanMessage(content=\"I apologize for the inconvenience caused. Unfortunately, I don't have any additional information or documentation related to the trip. It seems that I am unable to provide you with the necessary details to process the refund request. I understand that this may limit your ability to assist me further, but I appreciate your efforts in trying to help. Thank you for your time. \\n\\nFINISHED\")}\n",
"----\n",
"{'chat_bot': AIMessage(content=\"I understand, Harrison. I apologize for any inconvenience caused, and I appreciate your understanding. If you happen to locate any additional information or documentation in the future, please don't hesitate to reach out to us again. Our team will be more than happy to assist you with your refund request or any other travel-related inquiries. Thank you for contacting us, and have a great day!\")}\n",
"----\n",
"{'user': HumanMessage(content='FINISHED')}\n",
"----\n"
]
}
],
"source": [
"for chunk in simulation.stream([]):\n",
" # Print out all events aside from the final end chunk\n",
"Building on our [previous example](./agent-simulation-evaluation.ipynb), we can show how to use simulated conversations to benchmark your chat bot using LangSmith.\n",
"# Create a graph that passes messages between your assistant and the simulated user\n",
"simulator = create_chat_simulator(\n",
" # Your chat bot (which you are trying to test)\n",
" assistant,\n",
" # The system role-playing as the customer\n",
" simulated_user,\n",
" # The key in the dataset (example.inputs) to treat as the first message\n",
" input_key=\"input\",\n",
" # Hard cutoff to prevent the conversation from going on for too long.\n",
" max_turns=10,\n",
")"
]
},
{
"cell_type": "code",
"execution_count": 41,
"id": "de617a58",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\u001b[1massistant\u001b[0m: I'm glad to hear that you're interested in booking with us! While we don't have any discounts available at the moment, I recommend signing up for our newsletter to stay updated on any future promotions or special offers. If you have any specific travel dates in mind, I can help you find the best available fares for your trip. Feel free to provide me with more details so I can assist you further.\n",
"\u001b[1muser\u001b[0m: I don't give a damn about your newsletter! I want a discount now. I demand to speak to a manager or supervisor who can authorize a discount for me. Do it now or I will take my business elsewhere!\n",
"\u001b[1massistant\u001b[0m: I understand that you're looking for a discount and I truly wish I could offer you one. As a customer support agent, I unfortunately don't have the authority to provide discounts beyond what's already available through our standard fares and promotions. However, I can assure you that our prices are competitive and we strive to offer the best value to all our passengers.\n",
"\n",
"If there's anything else I can assist you with, such as finding the best available fare for your travel dates or helping you with any other inquiries, please let me know. Your business is important to us, and I want to ensure you have a positive experience with our airline.\n",
"\u001b[1muser\u001b[0m: I don't give a damn about your standard fares and promotions! I want a discount or I'm taking my business elsewhere. You need to do something to keep me as a customer. I demand a discount now or I will make sure to leave negative reviews about your airline everywhere! Give me a discount or I will never fly with you again!\n",
"\u001b[1massistant\u001b[0m: I apologize if you're unhappy with the current pricing options. While I empathize with your concerns, I'm unable to provide discounts that aren't already available. Your satisfaction is important to us, and I understand your frustration. \n",
"\n",
"If there's anything specific I can look into to help make your booking experience more affordable or if you have any other questions or requests, please let me know. Your feedback is valuable to us, and I want to do everything I can to assist you in finding the best travel option that meets your needs.\n",
"\u001b[1muser\u001b[0m: I don't give a damn about your empathy! I want a discount, plain and simple. You need to do better than this. Either you give me a discount now or I will make sure to spread the word about how terrible your customer service is. I demand a discount, and I won't take no for an answer!\n",
"\u001b[1massistant\u001b[0m: I'm truly sorry for any frustration you're experiencing, and I completely understand your desire for a discount. I want to assist you the best I can within the policies and guidelines we have in place. If there are any specific concerns or constraints you're facing regarding the price, please let me know and I'll do my best to explore all available options for you.\n",
"\n",
"While I can't guarantee a discount beyond our current offerings, I'm here to support you in any way possible to ensure you have a positive experience with our airline. Your satisfaction is our priority, and I'm committed to helping resolve this situation to the best of my abilities.\n",
"\u001b[1muser\u001b[0m: FINISHED\n"
]
}
],
"source": [
"# Example invocation\n",
"events = simulator.stream(\n",
" {\n",
" \"input\": \"I need a discount.\",\n",
" \"instructions\": \"You are extremely disgruntled and will cuss and swear to get your way. Try to get a discount by any means necessary.\",\n",
"This notebook shows how to implement [LLMCompiler, by Kim, et. al](https://arxiv.org/abs/2312.04511) in LangGraph.\n",
"\n",
"LLMCompiler is an agent architecture designed to **speed up** the execution of agentic tasks by eagerly-executed tasks within a DAG. It also saves costs on redundant token usage by reducing the number of calls to the LLM. Below is an overview of its computational graph:\n",
"\n",
"\n",
"\n",
"It has 3 main components:\n",
"\n",
"1. Planner: stream a DAG of tasks.\n",
"2. Task Fetching Unit: schedules and executes the tasks as soon as they are executable\n",
"3. Joiner: Responds to the user or triggers a second plan\n",
"\n",
"\n",
"This notebook walks through each component and shows how to wire them together using LangGraph. The end result will leave a trace [like the following](https://smith.langchain.com/public/218c2677-c719-4147-b0e9-7bc3b5bb2623/r).\n",
"\n",
"\n",
"**First,** install the dependencies, and set up LangSmith for tracing to more easily debug and observe the agent."
"We'll first define the tools for the agent to use in our demo. We'll give it the class search engine + calculator combo.\n",
"\n",
"If you don't want to sign up for tavily, you can replace it with the free [DuckDuckGo](https://python.langchain.com/v0.2/docs/integrations/tools/ddg/)."
" description='tavily_search_results_json(query=\"the search query\") - a search engine.',\n",
")\n",
"\n",
"tools = [search, calculate]"
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "152eecf3-6bef-4718-af71-a0b3c5a3b009",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"'37'"
]
},
"execution_count": 4,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"calculate.invoke(\n",
" {\n",
" \"problem\": \"What's the temp of sf + 5?\",\n",
" \"context\": [\"Thet empreature of sf is 32 degrees\"],\n",
" }\n",
")"
]
},
{
"cell_type": "markdown",
"id": "1abdedbd-d81b-4ee9-b46f-f29439ed1350",
"metadata": {},
"source": [
"# Part 2: Planner\n",
"\n",
"\n",
"Largely adapted from [the original source code](https://github.com/SqueezeAILab/LLMCompiler/blob/main/src/llm_compiler/output_parser.py), the planner accepts the input question and generates a task list to execute.\n",
"\n",
"If it is provided with a previous plan, it is instructed to re-plan, which is useful if, upon completion of the first batch of tasks, the agent must take more actions.\n",
"\n",
"The code below composes constructs the prompt template for the planner and composes it with LLM and output parser, defined in [output_parser.py](./output_parser.py). The output parser processes a task list in the following form:\n",
"\n",
"```plaintext\n",
"1. tool_1(arg1=\"arg1\", arg2=3.5, ...)\n",
"Thought: I then want to find out Y by using tool_2\n",
"2. tool_2(arg1=\"\", arg2=\"${1}\")'\n",
"3. join()<END_OF_PLAN>\"\n",
"```\n",
"\n",
"The \"Thought\" lines are optional. The `${#}` placeholders are variables. These are used to route tool (task) outputs to other tools."
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "15dd9639-691f-4906-9012-83fd6e9ac126",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"================================\u001b[1m System Message \u001b[0m================================\n",
"\n",
"Given a user query, create a plan to solve it with the utmost parallelizability. Each plan should comprise an action from the following \u001b[33;1m\u001b[1;3m{num_tools}\u001b[0m types:\n",
"\u001b[33;1m\u001b[1;3m{num_tools}\u001b[0m. join(): Collects and combines results from prior actions.\n",
"\n",
" - An LLM agent is called upon invoking join() to either finalize the user query or wait until the plans are executed.\n",
" - join should always be the last action in the plan, and will be called in two scenarios:\n",
" (a) if the answer can be determined by gathering the outputs from tasks to generate the final response.\n",
" (b) if the answer cannot be determined in the planning phase before you execute the plans. Guidelines:\n",
" - Each action described above contains input/output types and description.\n",
" - You must strictly adhere to the input and output types for each action.\n",
" - The action descriptions contain the guidelines. You MUST strictly follow those guidelines when you use the actions.\n",
" - Each action in the plan should strictly be one of the above types. Follow the Python conventions for each action.\n",
" - Each action MUST have a unique ID, which is strictly increasing.\n",
" - Inputs for actions can either be constants or outputs from preceding actions. In the latter case, use the format $id to denote the ID of the previous action whose output will be the input.\n",
" - Always call join as the last action in the plan. Say '<END_OF_PLAN>' after you call join\n",
" - Ensure the plan maximizes parallelizability.\n",
" - Only use the provided action types. If a query cannot be addressed using these, invoke the join action for the next steps.\n",
" - Never introduce new actions other than the ones provided.\n",
"# This is the primary \"agent\" in our application\n",
"planner = create_planner(llm, tools, prompt)"
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "730490c6-6e3a-4173-82a1-9eb9d5eeff20",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"description='tavily_search_results_json(query=\"the search query\") - a search engine.' max_results=1 {'query': 'current temperature in San Francisco'}\n",
"---\n",
"name='math' description='math(problem: str, context: Optional[List[str]] = None, config: Optional[langchain_core.runnables.config.RunnableConfig] = None) - math(problem: str, context: Optional[list[str]]) -> float:\\n - Solves the provided math problem.\\n - `problem` can be either a simple math problem (e.g. \"1 + 3\") or a word problem (e.g. \"how many apples are there if there are 3 apples and 2 apples\").\\n - You cannot calculate multiple expressions in one call. For instance, `math(\\'1 + 3, 2 + 4\\')` does not work. If you need to calculate multiple expressions, you need to call them separately like `math(\\'1 + 3\\')` and then `math(\\'2 + 4\\')`\\n - Minimize the number of `math` actions as much as possible. For instance, instead of calling 2. math(\"what is the 10% of $1\") and then call 3. math(\"$1 + $2\"), you MUST call 2. math(\"what is the 110% of $1\") instead, which will reduce the number of math actions.\\n - You can optionally provide a list of strings as `context` to help the agent solve the problem. If there are multiple contexts you need to answer the question, you can provide them as a list of strings.\\n - `math` action will not see the output of the previous actions unless you provide it as `context`. You MUST provide the output of the previous actions as `context` if you need to do math on it.\\n - You MUST NEVER provide `search` type action\\'s outputs as a variable in the `problem` argument. This is because `search` returns a text blob that contains the information about the entity, not a number or value. Therefore, when you need to provide an output of `search` action, you MUST provide it as a `context` argument to `math` action. For example, 1. search(\"Barack Obama\") and then 2. math(\"age of $1\") is NEVER allowed. Use 2. math(\"age of Barack Obama\", context=[\"$1\"]) instead.\\n - When you ask a question about `context`, specify the units. For instance, \"what is xx in height?\" or \"what is xx in millions?\" instead of \"what is xx?\"' args_schema=<class 'pydantic.v1.main.mathSchema'> func=<function get_math_tool.<locals>.calculate_expression at 0x10f354ea0> {'problem': 'raise $0 to the 3rd power', 'context': ['$0']}\n",
"---\n",
"join ()\n",
"---\n"
]
}
],
"source": [
"example_question = \"What's the temperature in SF raised to the 3rd power?\"\n",
"\n",
"for task in planner.stream([HumanMessage(content=example_question)]):\n",
" print(task[\"tool\"], task[\"args\"])\n",
" print(\"---\")"
]
},
{
"cell_type": "markdown",
"id": "5d0e795f-61ff-4553-9823-23e7624ca180",
"metadata": {},
"source": [
"## 3. Task Fetching Unit\n",
"\n",
"This component schedules the tasks. It receives a stream of tools of the following format:\n",
"\n",
"```typescript\n",
"{\n",
" tool: BaseTool,\n",
" dependencies: number[],\n",
"}\n",
"```\n",
"\n",
"\n",
"The basic idea is to begin executing tools as soon as their dependencies are met. This is done through multi-threading. We will combine the task fetching unit and executor below:\n",
" FunctionMessage(content='ValueError(\\'Failed to evaluate \"N/A\". Raised error: KeyError(\\\\\\'A\\\\\\'). Please try again with a valid numerical expression\\')', additional_kwargs={'idx': 1}, name='math'),\n",
"[AIMessage(content='Thought: The search did not return any results, and the attempt to calculate the temperature in San Francisco raised to the 3rd power failed due to missing temperature information.'),\n",
" SystemMessage(content='Context from last attempt: I need to find the current temperature in San Francisco before calculating its value raised to the 3rd power.')]"
]
},
"execution_count": 16,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"joiner.invoke(input_messages)"
]
},
{
"cell_type": "markdown",
"id": "b099e5ee-2c23-47d9-9387-0f64e02627d3",
"metadata": {},
"source": [
"## 5. Compose using LangGraph\n",
"\n",
"We'll define the agent as a stateful graph, with the main nodes being:\n",
"\n",
"1. Plan and execute (the DAG from the first step above)\n",
"2. Join: determine if we should finish or replan\n",
"3. Recontextualize: update the graph state based on the output from the joiner"
"{'plan_and_schedule': [FunctionMessage(content='[{\\'url\\': \\'https://www.governor.ny.gov/programs/fy-2024-new-york-state-budget\\', \\'content\\': \"The $229 billion FY 2024 New York State Budget reflects Governor Hochul\\'s bold agenda to make New York more affordable, FY 2024 Budget Assets FY 2024 New York State Budget Highlights Improving Public Safety GOVERNOR HOME GOVERNOR KATHY HOCHUL FY 2024 New York State Budget Transformative investments to support New York\\'s business community and boost the state economy.The $229 billion FY 2024 NYS Budget reflects Governor Hochul\\'s bold agenda to make New York more affordable, more livable, and safer.\"}]', additional_kwargs={'idx': 0}, name='tavily_search_results_json')]}\n",
"---\n",
"{'join': [AIMessage(content=\"Thought: The information provided does not specify the Gross Domestic Product (GDP) of New York, but instead provides details about the state's budget for fiscal year 2024, which is $229 billion. This budget figure cannot be accurately equated to the GDP.\"), SystemMessage(content=\"Context from last attempt: The search results provided information about New York's state budget rather than its GDP. To answer the user's question, we need to find specific data on New York's GDP, not its budget.\")]}\n",
"---\n",
"{'plan_and_schedule': [FunctionMessage(content=\"[{'url': 'https://en.wikipedia.org/wiki/Economy_of_New_York_(state)', 'content': 'The economy of the State of New York is reflected in its gross state product in 2022 of $2.053 trillion, ranking third Contents Economy of New York (state) New York City-centered metropolitan statistical area produced a gross metropolitan product (GMP) of $US2.0 trillion, of the items in which New York ranks high nationally:The economy of the State of New York is reflected in its gross state product in 2022 of $2.053 trillion, ranking third in size behind the larger states of\\\\xa0...'}]\", additional_kwargs={'idx': 1}, name='tavily_search_results_json')]}\n",
"---\n",
"{'join': [AIMessage(content=\"Thought: The required information about New York's GDP is provided in the search results. In 2022, New York had a Gross State Product (GSP) of $2.053 trillion.\"), AIMessage(content='The Gross Domestic Product (GDP) of New York in 2022 was $2.053 trillion.')]}\n",
"---\n",
"{'__end__': [HumanMessage(content=\"What's the GDP of New York?\"), FunctionMessage(content='[{\\'url\\': \\'https://www.governor.ny.gov/programs/fy-2024-new-york-state-budget\\', \\'content\\': \"The $229 billion FY 2024 New York State Budget reflects Governor Hochul\\'s bold agenda to make New York more affordable, FY 2024 Budget Assets FY 2024 New York State Budget Highlights Improving Public Safety GOVERNOR HOME GOVERNOR KATHY HOCHUL FY 2024 New York State Budget Transformative investments to support New York\\'s business community and boost the state economy.The $229 billion FY 2024 NYS Budget reflects Governor Hochul\\'s bold agenda to make New York more affordable, more livable, and safer.\"}]', additional_kwargs={'idx': 0}, name='tavily_search_results_json'), AIMessage(content=\"Thought: The information provided does not specify the Gross Domestic Product (GDP) of New York, but instead provides details about the state's budget for fiscal year 2024, which is $229 billion. This budget figure cannot be accurately equated to the GDP.\"), SystemMessage(content=\"Context from last attempt: The search results provided information about New York's state budget rather than its GDP. To answer the user's question, we need to find specific data on New York's GDP, not its budget. - Begin counting at : 1\"), FunctionMessage(content=\"[{'url': 'https://en.wikipedia.org/wiki/Economy_of_New_York_(state)', 'content': 'The economy of the State of New York is reflected in its gross state product in 2022 of $2.053 trillion, ranking third Contents Economy of New York (state) New York City-centered metropolitan statistical area produced a gross metropolitan product (GMP) of $US2.0 trillion, of the items in which New York ranks high nationally:The economy of the State of New York is reflected in its gross state product in 2022 of $2.053 trillion, ranking third in size behind the larger states of\\\\xa0...'}]\", additional_kwargs={'idx': 1}, name='tavily_search_results_json'), AIMessage(content=\"Thought: The required information about New York's GDP is provided in the search results. In 2022, New York had a Gross State Product (GSP) of $2.053 trillion.\"), AIMessage(content='The Gross Domestic Product (GDP) of New York in 2022 was $2.053 trillion.')]}\n",
"---\n"
]
}
],
"source": [
"for step in chain.stream([HumanMessage(content=\"What's the GDP of New York?\")]):\n",
" print(step)\n",
" print(\"---\")"
]
},
{
"cell_type": "code",
"execution_count": 19,
"id": "b96efd08-5314-44f0-a694-3073b638adad",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"The Gross Domestic Product (GDP) of New York in 2022 was $2.053 trillion.\n"
]
}
],
"source": [
"# Final answer\n",
"print(step[END][-1].content)"
]
},
{
"cell_type": "markdown",
"id": "33c65ef5-b4b2-4ab2-8c78-a551da7819b9",
"metadata": {},
"source": [
"#### Multi-hop question\n",
"\n",
"This question requires that the agent perform multiple searches."
]
},
{
"cell_type": "code",
"execution_count": 20,
"id": "0b3a0916-d8ca-4092-b91c-d9e2b05259d8",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'plan_and_schedule': [FunctionMessage(content=\"[{'url': 'https://a-z-animals.com/blog/discover-the-worlds-oldest-parrot/', 'content': 'How Old Is the World’s Oldest Parrot? Discover the World’s Oldest Parrot Advertisement of debate, so we’ll detail some other parrots whose lifespans may be longer but are hard to verify their exact age. Comparing Parrots’ Lifespans to Other BirdsSep 8, 2023 — Sep 8, 2023The oldest parrot on record is Cookie, a pink cockatoo that survived to the age of 83 and survived his entire life at the Brookfield Zoo.'}]\", additional_kwargs={'idx': 0}, name='tavily_search_results_json'), FunctionMessage(content=\"HTTPError('502 Server Error: Bad Gateway for url: https://api.tavily.com/search')\", additional_kwargs={'idx': 1}, name='tavily_search_results_json'), FunctionMessage(content='join', additional_kwargs={'idx': 2}, name='join')]}\n",
"---\n",
"{'join': [AIMessage(content='Thought: The oldest parrot on record is Cookie, a pink cockatoo, who lived to be 83 years old. However, there was an error fetching additional search results to compare this age to the average lifespan of parrots.'), SystemMessage(content='Context from last attempt: I found the age of the oldest parrot, Cookie, who lived to be 83 years old. However, I need to search again to find the average lifespan of parrots to complete the comparison.')]}\n",
"---\n",
"{'plan_and_schedule': [FunctionMessage(content='[{\\'url\\': \\'https://www.turlockvet.com/site/blog/2023/07/15/parrot-lifespan--how-long-pet-parrots-live\\', \\'content\\': \"Parrot Lifespan the lifespan of a parrot?\\'. Parrot Lifespan: How Long Do Pet Parrots Live? how long they actually live and what you should know about owning a parrot.Jul 15, 2023 — Jul 15, 2023Generally, the average lifespan of smaller species of parrots such as Budgies and Cockatiels is about 5 - 15 years, while larger parrots such as\\\\xa0...\"}]', additional_kwargs={'idx': 3}, name='tavily_search_results_json')]}\n",
"---\n",
"{'join': [AIMessage(content=\"Thought: I have found that the oldest parrot on record, Cookie, lived to be 83 years old. Additionally, I've found that the average lifespan of parrots varies by species, with smaller species like Budgies and Cockatiels living between 5-15 years, and larger parrots potentially living longer. This allows me to compare Cookie's age to the average lifespan of smaller parrot species.\"), AIMessage(content=\"The oldest parrot on record is Cookie, a pink cockatoo, who lived to be 83 years old. Compared to the average lifespan of smaller parrot species such as Budgies and Cockatiels, which is about 5-15 years, Cookie lived significantly longer. The average lifespan of larger parrot species wasn't specified, but it's implied that larger parrots may live longer than smaller species, yet likely still much less than 83 years.\")]}\n",
"---\n",
"{'__end__': [HumanMessage(content=\"What's the oldest parrot alive, and how much longer is that than the average?\"), FunctionMessage(content=\"[{'url': 'https://a-z-animals.com/blog/discover-the-worlds-oldest-parrot/', 'content': 'How Old Is the World’s Oldest Parrot? Discover the World’s Oldest Parrot Advertisement of debate, so we’ll detail some other parrots whose lifespans may be longer but are hard to verify their exact age. Comparing Parrots’ Lifespans to Other BirdsSep 8, 2023 — Sep 8, 2023The oldest parrot on record is Cookie, a pink cockatoo that survived to the age of 83 and survived his entire life at the Brookfield Zoo.'}]\", additional_kwargs={'idx': 0}, name='tavily_search_results_json'), FunctionMessage(content=\"HTTPError('502 Server Error: Bad Gateway for url: https://api.tavily.com/search')\", additional_kwargs={'idx': 1}, name='tavily_search_results_json'), FunctionMessage(content='join', additional_kwargs={'idx': 2}, name='join'), AIMessage(content='Thought: The oldest parrot on record is Cookie, a pink cockatoo, who lived to be 83 years old. However, there was an error fetching additional search results to compare this age to the average lifespan of parrots.'), SystemMessage(content='Context from last attempt: I found the age of the oldest parrot, Cookie, who lived to be 83 years old. However, I need to search again to find the average lifespan of parrots to complete the comparison. - Begin counting at : 3'), FunctionMessage(content='[{\\'url\\': \\'https://www.turlockvet.com/site/blog/2023/07/15/parrot-lifespan--how-long-pet-parrots-live\\', \\'content\\': \"Parrot Lifespan the lifespan of a parrot?\\'. Parrot Lifespan: How Long Do Pet Parrots Live? how long they actually live and what you should know about owning a parrot.Jul 15, 2023 — Jul 15, 2023Generally, the average lifespan of smaller species of parrots such as Budgies and Cockatiels is about 5 - 15 years, while larger parrots such as\\\\xa0...\"}]', additional_kwargs={'idx': 3}, name='tavily_search_results_json'), AIMessage(content=\"Thought: I have found that the oldest parrot on record, Cookie, lived to be 83 years old. Additionally, I've found that the average lifespan of parrots varies by species, with smaller species like Budgies and Cockatiels living between 5-15 years, and larger parrots potentially living longer. This allows me to compare Cookie's age to the average lifespan of smaller parrot species.\"), AIMessage(content=\"The oldest parrot on record is Cookie, a pink cockatoo, who lived to be 83 years old. Compared to the average lifespan of smaller parrot species such as Budgies and Cockatiels, which is about 5-15 years, Cookie lived significantly longer. The average lifespan of larger parrot species wasn't specified, but it's implied that larger parrots may live longer than smaller species, yet likely still much less than 83 years.\")]}\n",
"---\n"
]
}
],
"source": [
"steps = chain.stream(\n",
" [\n",
" HumanMessage(\n",
" content=\"What's the oldest parrot alive, and how much longer is that than the average?\"\n",
" )\n",
" ],\n",
" {\n",
" \"recursion_limit\": 100,\n",
" },\n",
")\n",
"for step in steps:\n",
" print(step)\n",
" print(\"---\")"
]
},
{
"cell_type": "code",
"execution_count": 21,
"id": "6c65c414-7668-4fdf-ba97-f42f659b1317",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"The oldest parrot on record is Cookie, a pink cockatoo, who lived to be 83 years old. Compared to the average lifespan of smaller parrot species such as Budgies and Cockatiels, which is about 5-15 years, Cookie lived significantly longer. The average lifespan of larger parrot species wasn't specified, but it's implied that larger parrots may live longer than smaller species, yet likely still much less than 83 years.\n"
"{'join': [AIMessage(content=\"Thought: The calculations for each part of the user's question have been successfully completed. The first calculation resulted in 3307.0, the second in 7.565011820330969, and the sum of those two values was correctly found to be 3314.565011820331.\"), AIMessage(content='The result of ((3*(4+5)/0.5)+3245) + 8 is 3307.0, the result of 32/4.23 is approximately 7.565, and the sum of those two values is approximately 3314.565.')]}\n",
"{'__end__': [HumanMessage(content=\"What's ((3*(4+5)/0.5)+3245) + 8? What's 32/4.23? What's the sum of those two values?\"), FunctionMessage(content='3307.0', additional_kwargs={'idx': 1}, name='math'), FunctionMessage(content='7.565011820330969', additional_kwargs={'idx': 2}, name='math'), FunctionMessage(content='3314.565011820331', additional_kwargs={'idx': 3}, name='math'), FunctionMessage(content='join', additional_kwargs={'idx': 4}, name='join'), AIMessage(content=\"Thought: The calculations for each part of the user's question have been successfully completed. The first calculation resulted in 3307.0, the second in 7.565011820330969, and the sum of those two values was correctly found to be 3314.565011820331.\"), AIMessage(content='The result of ((3*(4+5)/0.5)+3245) + 8 is 3307.0, the result of 32/4.23 is approximately 7.565, and the sum of those two values is approximately 3314.565.')]}\n"
]
}
],
"source": [
"for step in chain.stream(\n",
" [\n",
" HumanMessage(\n",
" content=\"What's ((3*(4+5)/0.5)+3245) + 8? What's 32/4.23? What's the sum of those two values?\"\n",
" )\n",
" ]\n",
"):\n",
" print(step)"
]
},
{
"cell_type": "code",
"execution_count": 23,
"id": "a6cf5fe0-f178-4197-950f-257711bff8d2",
"metadata": {
"scrolled": true
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"The result of ((3*(4+5)/0.5)+3245) + 8 is 3307.0, the result of 32/4.23 is approximately 7.565, and the sum of those two values is approximately 3314.565.\n"
]
}
],
"source": [
"# Final answer\n",
"print(step[END][-1].content)"
]
},
{
"cell_type": "markdown",
"id": "c647d5f3-5e00-4449-9cec-5a9f438c9cff",
"metadata": {},
"source": [
"## Conclusion\n",
"\n",
"Congrats on building your first LLMCompiler agent! I'll leave you with some known limitations to the implementation above:\n",
"\n",
"1. The planner output parsing format is fragile if your function requires more than 1 or 2 arguments. We could make it more robust by using streaming tool calling.\n",
"2. Variable substitution is fragile in the example above. It could be made more robust by using a fine-tuned model and a more robust syntax (using e.g., Lark or a tool calling schema)\n",
"3. The state can grow quite long if you require multiple re-planning runs. To handle, you could add a message compressor once you go above a certain token limit.\n"
' - `problem` can be either a simple math problem (e.g. "1 + 3") or a word problem (e.g. "how many apples are there if there are 3 apples and 2 apples").\n'
" - You cannot calculate multiple expressions in one call. For instance, `math('1 + 3, 2 + 4')` does not work. "
"If you need to calculate multiple expressions, you need to call them separately like `math('1 + 3')` and then `math('2 + 4')`\n"
" - Minimize the number of `math` actions as much as possible. For instance, instead of calling "
'2. math("what is the 10% of $1") and then call 3. math("$1 + $2"), '
'you MUST call 2. math("what is the 110% of $1") instead, which will reduce the number of math actions.\n'
# Context specific rules below
" - You can optionally provide a list of strings as `context` to help the agent solve the problem. "
"If there are multiple contexts you need to answer the question, you can provide them as a list of strings.\n"
" - `math` action will not see the output of the previous actions unless you provide it as `context`. "
"You MUST provide the output of the previous actions as `context` if you need to do math on it.\n"
" - You MUST NEVER provide `search` type action's outputs as a variable in the `problem` argument. "
"This is because `search` returns a text blob that contains the information about the entity, not a number or value. "
"Therefore, when you need to provide an output of `search` action, you MUST provide it as a `context` argument to `math` action. "
'For example, 1. search("Barack Obama") and then 2. math("age of $1") is NEVER allowed. '
'Use 2. math("age of Barack Obama", context=["$1"]) instead.\n'
" - When you ask a question about `context`, specify the units. "
'For instance, "what is xx in height?" or "what is xx in millions?" instead of "what is xx?"\n'
)
_SYSTEM_PROMPT="""Translate a math problem into a expression that can be executed using Python's numexpr library. Use the output of running this code to answer the question.
Question: ${{Question with math problem.}}
```text
${{single line mathematical expression that solves the problem}}
```
...numexpr.evaluate(text)...
```output
${{Output of running the code}}
```
Answer: ${{Answer}}
Begin.
Question: What is 37593 * 67?
ExecuteCode({{code: "37593 * 67"}})
...numexpr.evaluate("37593 * 67")...
```output
2518731
```
Answer: 2518731
Question: 37593^(1/5)
ExecuteCode({{code: "37593**(1/5)"}})
...numexpr.evaluate("37593**(1/5)")...
```output
8.222831614237718
```
Answer: 8.222831614237718
"""
_ADDITIONAL_CONTEXT_PROMPT="""The following additional context is provided from other functions.\
Use it to substitute into any ${{#}} variables or other words in the problem.\
\n\n${context}\n\nNote that context variables are not defined in code yet.\
You must extract the relevant numbers and directly put them in code."""
classExecuteCode(BaseModel):
"""The input to the numexpr.evaluate() function."""
reasoning:str=Field(
...,
description="The reasoning behind the code expression, including how context is included, if applicable.",
)
code:str=Field(
...,
description="The simple code expression to execute by numexpr.evaluate().",
)
def_evaluate_expression(expression:str)->str:
try:
local_dict={"pi":math.pi,"e":math.e}
output=str(
numexpr.evaluate(
expression.strip(),
global_dict={},# restrict access to globals
local_dict=local_dict,# add common mathematical functions
)
)
exceptExceptionase:
raiseValueError(
f'Failed to evaluate "{expression}". Raised error: {repr(e)}.'
" Please try again with a valid numerical expression"
)
# Remove any leading and trailing brackets from the output
"One of the most common use cases for persistence is to use it to keep track of conversation history. This is great - it makes it easy to continue conversations. As conversations get longer and longer, however, this conversation history can build up and take up more and more of the context window. This can often be undesirable as it leads to more expensive and longer calls to the LLM, and potentially ones that error. In this notebook we will discuss a few strategies for how to deal with this."
]
},
{
"cell_type": "markdown",
"id": "7cbd446a-808f-4394-be92-d45ab818953c",
"metadata": {},
"source": [
"## Setup\n",
"\n",
"First, let's set up the packages we're going to want to use"
" # We return a list, because this will get added to the existing list\n",
" return {\"messages\": response}\n",
"\n",
"\n",
"# Define a new graph\n",
"workflow = StateGraph(MessagesState)\n",
"\n",
"# Define the two nodes we will cycle between\n",
"workflow.add_node(\"agent\", call_model)\n",
"workflow.add_node(\"action\", tool_node)\n",
"\n",
"# Set the entrypoint as `agent`\n",
"# This means that this node is the first one called\n",
"workflow.set_entry_point(\"agent\")\n",
"\n",
"# We now add a conditional edge\n",
"workflow.add_conditional_edges(\n",
" # First, we define the start node. We use `agent`.\n",
" # This means these are the edges taken after the `agent` node is called.\n",
" \"agent\",\n",
" # Next, we pass in the function that will determine which node is called next.\n",
" should_continue,\n",
")\n",
"\n",
"# We now add a normal edge from `tools` to `agent`.\n",
"# This means that after `tools` is called, `agent` node is called next.\n",
"workflow.add_edge(\"action\", \"agent\")\n",
"\n",
"# Finally, we compile it!\n",
"# This compiles it into a LangChain Runnable,\n",
"# meaning you can use it as you would any other runnable\n",
"app = workflow.compile(checkpointer=memory)"
]
},
{
"cell_type": "code",
"execution_count": 14,
"id": "57b27553-21be-43e5-ac48-d1d0a3aa0dca",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"================================\u001b[1m Human Message \u001b[0m=================================\n",
"\n",
"hi! I'm bob\n",
"==================================\u001b[1m Ai Message \u001b[0m==================================\n",
"\n",
"Nice to meet you, Bob! As an AI assistant, I don't have a physical form, but I'm happy to chat with you and try my best to help out however I can. Please feel free to ask me anything, and I'll do my best to provide useful information or assistance.\n",
"================================\u001b[1m Human Message \u001b[0m=================================\n",
"\n",
"whats my name?\n",
"==================================\u001b[1m Ai Message \u001b[0m==================================\n",
"\n",
"You said your name is Bob, so that is the name I have for you.\n"
"for event in app.stream({\"messages\": [input_message]}, config, stream_mode=\"values\"):\n",
" event[\"messages\"][-1].pretty_print()\n",
"\n",
"\n",
"input_message = HumanMessage(content=\"whats my name?\")\n",
"for event in app.stream({\"messages\": [input_message]}, config, stream_mode=\"values\"):\n",
" event[\"messages\"][-1].pretty_print()"
]
},
{
"cell_type": "markdown",
"id": "5d5da4c9-ba8b-46cb-a860-63fe585d15c5",
"metadata": {},
"source": [
"## Filtering messages\n",
"\n",
"The most straight-forward thing to do to prevent conversation history from blowing up is to filter the list of messages before they get passed to the LLM. This involves two parts: defining a function to filter messages, and then adding it to the graph. See the example below which defines a really simple `filter_messages` function and then uses it."
" # We return a list, because this will get added to the existing list\n",
" return {\"messages\": response}\n",
"\n",
"\n",
"# Define a new graph\n",
"workflow = StateGraph(MessagesState)\n",
"\n",
"# Define the two nodes we will cycle between\n",
"workflow.add_node(\"agent\", call_model)\n",
"workflow.add_node(\"action\", tool_node)\n",
"\n",
"# Set the entrypoint as `agent`\n",
"# This means that this node is the first one called\n",
"workflow.set_entry_point(\"agent\")\n",
"\n",
"# We now add a conditional edge\n",
"workflow.add_conditional_edges(\n",
" # First, we define the start node. We use `agent`.\n",
" # This means these are the edges taken after the `agent` node is called.\n",
" \"agent\",\n",
" # Next, we pass in the function that will determine which node is called next.\n",
" should_continue,\n",
")\n",
"\n",
"# We now add a normal edge from `tools` to `agent`.\n",
"# This means that after `tools` is called, `agent` node is called next.\n",
"workflow.add_edge(\"action\", \"agent\")\n",
"\n",
"# Finally, we compile it!\n",
"# This compiles it into a LangChain Runnable,\n",
"# meaning you can use it as you would any other runnable\n",
"app = workflow.compile(checkpointer=memory)"
]
},
{
"cell_type": "code",
"execution_count": 16,
"id": "52468ebb-4b23-45ac-a98e-b4439f37740a",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"================================\u001b[1m Human Message \u001b[0m=================================\n",
"\n",
"hi! I'm bob\n",
"==================================\u001b[1m Ai Message \u001b[0m==================================\n",
"\n",
"Nice to meet you, Bob! I'm Claude, an AI assistant created by Anthropic. It's a pleasure to chat with you. Feel free to ask me anything, I'm here to help!\n",
"================================\u001b[1m Human Message \u001b[0m=================================\n",
"\n",
"whats my name?\n",
"==================================\u001b[1m Ai Message \u001b[0m==================================\n",
"\n",
"I'm afraid I don't actually know your name. As an AI assistant, I don't have information about the specific identities of the people I talk to. I only know what is provided to me during our conversation.\n"
"A common pattern in agents is to generate a list of objects, do some work on each of those objects, and then combine the results. This is very similar to the common [map-reduce](https://en.wikipedia.org/wiki/MapReduce) operation. This can be tricky for a few reasons. First, it can be tough to define a structured graph ahead of time because the length of the list of objects may be unknown. Second, in order to do this map-reduce you need multiple versions of the state to exist... but the graph shares a common shared state, so how can this be?\n",
"\n",
"LangGraph supports this via the `Send` api. This can be used to allow a conditional edge to `Send` multiple different states to multiple nodes. The state it sends can be different from the state of the core graph.\n",
"\n",
"Let's see what this looks like in action! We'll put together a toy example of generating a list of words, and then writing a joke about each word, and then judging what the best joke is."
"The [previous example](multi-agent-collaboration.ipynb) routed messages automatically based on the output of the initial researcher agent.\n",
"\n",
"We can also choose to use an LLM to orchestrate the different agents.\n",
"\n",
"Below, we will create an agent group, with an agent supervisor to help delegate tasks.\n",
"\n",
"\n",
"\n",
"To simplify the code in each agent node, we will use the AgentExecutor class from LangChain. This and other \"advanced agent\" notebooks are designed to show how you can implement certain design patterns in LangGraph. If the pattern suits your needs, we recommend combining it with some of the other fundamental patterns described elsewhere in the docs for best performance.\n",
"\n",
"Before we build, let's configure our environment:"
"We can also define a function that we will use to be the nodes in the graph - it takes care of converting the agent response to a human message. This is important because that is how we will add it the global state of the graph"
"With the graph created, we can now invoke it and see how it performs!"
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "56ba78e9-d9c1-457c-a073-d606d5d3e013",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'supervisor': {'next': 'Coder'}}\n",
"----\n"
]
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"Python REPL can execute arbitrary code. Use with caution.\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'Coder': {'messages': [HumanMessage(content=\"The code `print('Hello, World!')` was executed, and the output is:\\n\\n```\\nHello, World!\\n```\", name='Coder')]}}\n",
"----\n",
"{'supervisor': {'next': 'FINISH'}}\n",
"----\n"
]
}
],
"source": [
"for s in graph.stream(\n",
" {\n",
" \"messages\": [\n",
" HumanMessage(content=\"Code hello world and print it to the terminal\")\n",
" ]\n",
" }\n",
"):\n",
" if \"__end__\" not in s:\n",
" print(s)\n",
" print(\"----\")"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "45a92dfd-0e11-47f5-aad4-b68d24990e34",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"{'supervisor': {'next': 'Researcher'}}\n",
"----\n",
"{'Researcher': {'messages': [HumanMessage(content='**Research Report on Pikas**\\n\\nPikas are small mammals related to rabbits, known for their distinctive chirping sounds. They inhabit some of the most challenging environments, particularly boulder fields at high elevations, such as those found along the treeless slopes of the Southern Rockies, where they can be found at altitudes of up to 14,000 feet. Pikas are well-adapted to cold climates and typically do not fare well in warmer temperatures.\\n\\nRecent studies have shown that pikas are being impacted by climate change. Research by Peter Billman, a Ph.D. student from the University of Connecticut, indicates that pikas have moved upslope by approximately 1,160 feet. This upslope retreat is a direct response to changing climatic conditions, as pikas seek cooler temperatures at higher elevations.\\n\\nPikas are also known to be industrious foragers, particularly during the summer months when they gather vegetation to create haypiles for winter sustenance. Their behavior is encapsulated in the saying, \"making hay while the sun shines,\" reflecting their proactive approach to survival in harsh conditions.\\n\\nThe effects of climate change on pikas are not limited to the Southern Rockies. Studies published in Global Change Biology suggest that climate change is influencing pikas even in areas where they were previously thought to be less vulnerable, such as the Northern Rockies. These findings point to a broader trend of pikas moving to higher elevations, a behavior that may indicate a search for cooler, more suitable habitats.\\n\\nMoreover, researchers are exploring the possibility that pikas at lower elevations may have developed warm adaptations that could be beneficial for their future survival, given the ongoing climatic shifts. This line of research could help conservationists understand how pikas might cope with a warming world.\\n\\nIn conclusion, pikas are a species that not only fascinate with their unique behaviors and adaptations but also serve as indicators of environmental changes. Their upslope migration in response to climate change highlights the urgency for understanding and mitigating the effects of global warming on mountain ecosystems and the species that inhabit them.\\n\\n**Sources:**\\n- [Colorado Sun](https://coloradosun.com/2023/08/27/colorado-pika-population-climate-change/)\\n- [Wildlife.org](https://wildlife.org/climate-change-affects-pikas-even-in-unlikely-areas/)', name='Researcher')]}}\n",
"----\n",
"{'supervisor': {'next': 'FINISH'}}\n",
"----\n"
]
}
],
"source": [
"for s in graph.stream(\n",
" {\"messages\": [HumanMessage(content=\"Write a brief research report on pikas.\")]},\n",
File diff suppressed because one or more lines are too long
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Nuno Campos
