langgraph/docs/docs/concepts/low_level.md

Low Level Conceptual Guide

Graphs

At its core, LangGraph models agent workflows as graphs. 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: Python functions 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. To emphasize: Nodes and Edges are nothing more than Python functions - they can contain an LLM or just good ol' Python code.

In short: nodes do the work. edges tell what to do next.

LangGraph's underlying graph algorithm uses message passing to define a general program. When a Node completes its operation, it sends messages along one or more edges to other node(s). These recipient nodes then execute their functions, pass the resulting messages to the next set of nodes, and the process continues. Inspired by Google's Pregel system, the program proceeds in discrete "super-steps."

A super-step can be considered a single iteration over the graph nodes. Nodes that run in parallel are part of the same super-step, while nodes that run sequentially belong to separate super-steps. At the start of graph execution, all nodes begin in an inactive state. A node becomes active when it receives a new message (state) on any of its incoming edges (or "channels"). The active node then runs its function and responds with updates. At the end of each super-step, nodes with no incoming messages vote to halt by marking themselves as inactive. The graph execution terminates when all nodes are inactive and no messages are in transit.

StateGraph

The StateGraph class is the main graph class to uses. This is parameterized by a user defined State object.

MessageGraph

The MessageGraph class is a special type of graph. The State of a MessageGraph is ONLY a list of messages. This class is rarely used except for chatbots, as most applications require the State to be more complex than a list of messages.

Compiling your graph

To build your graph, you first define the state, you then add nodes and edges, and then you compile it. What exactly is compiling your graph and why is it needed?

Compiling is a pretty simple step. It provides a few basic checks on the structure of your graph (no orphaned nodes, etc). It is also where you can specify runtime args like checkpointers and breakpoints. You compile your graph by just calling the .compile method:

graph = graph_builder.compile(...)

You MUST compile your graph before you can use it.

State

The first thing you do when you define a graph is define the State of the graph. The State consists of the schema of the graph as well as reducer functions which specify how to apply updates to the state. The schema of the State will be the input schema to all Nodes and Edges in the graph, and can be either a TypedDict or a Pydantic model. All Nodes will emit updates to the State which are then applied using the specified reducer function.

Schema

The main documented way to specify the schema of a graph is by using TypedDict. However, we also support using a Pydantic BaseModel as your graph state to add default values and additional data validation.

By default, the graph will have the same input and output schemas. If you want to change this, you can also specify explicit input and output schemas directly. This is useful when you have a lot of keys, and some are explicitly for input and others for output. See the notebook here for how to use.

By default, all nodes in the graph will share the same state. This means that they will read and write to the same state channels. It is possible to have nodes write to private state channels inside the graph for internal node communication - see this notebook for how to do that.

Reducers

Reducers are key to understanding how updates from nodes are applied to the State. Each key in the State has its own independent reducer function. If no reducer function is explicitly specified then it is assumed that all updates to that key should override it. There are a few different types of reducers, starting with the default type of reducer:

Default Reducer

These two examples show how to use the default reducer:

Example A:

from typing import TypedDict

class State(TypedDict):
    foo: int
    bar: list[str]

In this example, no reducer functions are specified for any key. Let's assume the input to the graph is {"foo": 1, "bar": ["hi"]}. Let's then assume the first Node returns {"foo": 2}. This is treated as an update to the state. Notice that the Node does not need to return the whole State schema - just an update. After applying this update, the State would then be {"foo": 2, "bar": ["hi"]}. If the second node returns {"bar": ["bye"]} then the State would then be {"foo": 2, "bar": ["bye"]}

Example B:

from typing import TypedDict, Annotated
from operator import add

class State(TypedDict):
    foo: int
    bar: Annotated[list[str], add]

In this example, we've used the Annotated type to specify a reducer function (operator.add) for the second key (bar). Note that the first key remains unchanged. Let's assume the input to the graph is {"foo": 1, "bar": ["hi"]}. Let's then assume the first Node returns {"foo": 2}. This is treated as an update to the state. Notice that the Node does not need to return the whole State schema - just an update. After applying this update, the State would then be {"foo": 2, "bar": ["hi"]}. If the second node returns {"bar": ["bye"]} then the State would then be {"foo": 2, "bar": ["hi", "bye"]}. Notice here that the bar key is updated by adding the two lists together.

Context Reducer

You can use Context channels to define shared resources (such as database connections) that are managed outside of your graph's nodes and excluded from checkpointing. The context manager provided to the Context channel is entered before the first step of the graph execution and exited after the last step, allowing you to set up and clean up resources for the duration of the graph invocation. Read this how to to see an example of using the Context channel in your graph.

Working with Messages in Graph State

Why use messages?

Most modern LLM providers have a chat model interface that accepts a list of messages as input. LangChain's ChatModel in particular accepts a list of Message objects as inputs. These messages come in a variety of forms such as HumanMessage (user input) or AIMessage (LLM response). To read more about what message objects are, please refer to this conceptual guide.

Using Messages in your Graph

In many cases, it is helpful to store prior conversation history as a list of messages in your graph state. To do so, we can add a key (channel) to the graph state that stores a list of Message objects and annotate it with a reducer function (see messages key in the example below). The reducer function is vital to telling the graph how to update the list of Message objects in the state with each state update (for example, when a node sends an update). If you don't specify a reducer, every state update will overwrite the list of messages with the most recently provided value. If you wanted to simply append messages to the existing list, you could use operator.add as a reducer.

However, you might also want to manually update messages in your graph state (e.g. human-in-the-loop). If you were to use operator.add, the manual state updates you send to the graph would be appended to the existing list of messages, instead of updating existing messages. To avoid that, you need a reducer that can keep track of message IDs and overwrite existing messages, if updated. To achieve this, you can use the prebuilt add_messages function. For brand new messages, it will simply append to existing list, but it will also handle the updates for existing messages correctly.

Serialization

In addition to keeping track of message IDs, the add_messages function will also try to deserialize messages into LangChain Message objects whenever a state update is received on the messages channel. See more information on LangChain serialization/deserialization here. This allows sending graph inputs / state updates in the following format:

# this is supported
{"messages": [HumanMessage(content="message")]}

# and this is also supported
{"messages": [{"type": "human", "content": "message"}]}

Since the state updates are always deserialized into LangChain Messages when using add_messages, you should use dot notation to access message attributes, like state["messages"][-1].content. Below is an example of a graph that uses add_messages as it's reducer function.

from langchain_core.messages import AnyMessage
from langgraph.graph.message import add_messages
from typing import Annotated, TypedDict

class GraphState(TypedDict):
    messages: Annotated[list[AnyMessage], add_messages]

MessagesState

Since having a list of messages in your state is so common, there exists a prebuilt state called MessagesState which makes it easy to use messages. MessagesState is defined with a single messages key which is a list of AnyMessage objects and uses the add_messages reducer. Typically, there is more state to track than just messages, so we see people subclass this state and add more fields, like:

from langgraph.graph import MessagesState

class State(MessagesState):
    documents: list[str]

Nodes

In LangGraph, nodes are typically python functions (sync or async) where the first positional argument is the state, and (optionally), the second positional argument is a "config", containing optional configurable parameters (such as a thread_id).

Similar to NetworkX, you add these nodes to a graph using the [add_node][langgraph.graph.StateGraph.add_node] method:

from langchain_core.runnables import RunnableConfig
from langgraph.graph import StateGraph

builder = StateGraph(dict)


def my_node(state: dict, config: RunnableConfig):
    print("In node: ", config["configurable"]["user_id"])
    return {"results": f"Hello, {state['input']}!"}


# The second argument is optional
def my_other_node(state: dict):
    return state


builder.add_node("my_node", my_node)
builder.add_node("other_node", my_other_node)
...

Behind the scenes, functions are converted to RunnableLambda's, which add batch and async support to your function, along with native tracing and debugging.

If you add a node to graph without specifying a name, it will be given a default name equivalent to the function name.

builder.add_node(my_node)
# You can then create edges to/from this node by referencing it as `"my_node"`

START Node

The START Node is a special node that represents the node sends user input to the graph. The main purpose for referencing this node is to determine which nodes should be called first.

from langgraph.graph import START

graph.add_edge(START, "node_a")

END Node

The END Node is a special node that represents a terminal node. This node is referenced when you want to denote which edges have no actions after they are done.

from langgraph.graph import END

graph.add_edge("node_a", END)

Edges

Edges define how the logic is routed and how the graph decides to stop. This is a big part of how your agents work and how different nodes communicate with each other. There are a few key types of edges:

• Normal Edges: Go directly from one node to the next.
• Conditional Edges: Call a function to determine which node(s) to go to next.
• Entry Point: Which node to call first when user input arrives.
• Conditional Entry Point: Call a function to determine which node(s) to call first when user input arrives.

A node can have MULTIPLE outgoing edges. 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.

Normal Edges

If you always want to go from node A to node B, you can use the [add_edge][langgraph.graph.StateGraph.add_edge] method directly.

graph.add_edge("node_a", "node_b")

Conditional Edges

If you want to optionally route to 1 or more edges (or optionally terminate), you can use the [add_conditional_edges][langgraph.graph.StateGraph.add_conditional_edges] method. This method accepts the name of a node and a "routing function" to call after that node is executed:

graph.add_conditional_edges("node_a", routing_function)

Similar to nodes, the routing_function accept the current state of the graph and return a value.

By default, the return value routing_function is used as the name of the node (or a list of nodes) to send the state to next. All those nodes will be run in parallel as a part of the next superstep.

You can optionally provide a dictionary that maps the routing_function's output to the name of the next node.

graph.add_conditional_edges("node_a", routing_function, {True: "node_b", False: "node_c"})

Entry Point

The entry point is the first node(s) that are run when the graph starts. You can use the [add_edge][langgraph.graph.StateGraph.add_edge] method from the virtual [START][start] node to the first node to execute to specify where to enter the graph.

from langgraph.graph import START

graph.add_edge(START, "node_a")

Conditional Entry Point

A conditional entry point lets you start at different nodes depending on custom logic. You can use [add_conditional_edges][langgraph.graph.StateGraph.add_conditional_edges] from the virtual [START][start] node to accomplish this.

from langgraph.graph import START

graph.add_conditional_edges(START, routing_function)

You can optionally provide a dictionary that maps the routing_function's output to the name of the next node.

graph.add_conditional_edges(START, routing_function, {True: "node_b", False: "node_c"})

Send

By default, Nodes and Edges are defined ahead of time and operate on the same shared state. However, there can be cases where the exact edges are not known ahead of time and/or you may want different versions of State to exist at the same time. A common of example of this is with map-reduce design patterns. In this design pattern, a first node may generate a list of objects, and you may want to apply some other node to all those objects. The number of objects may be unknown ahead of time (meaning the number of edges may not be known) and the input State to the downstream Node should be different (one for each generated object).

To support this design pattern, LangGraph supports returning Send objects from conditional edges. Send takes two arguments: first is the name of the node, and second is the state to pass to that node.

def continue_to_jokes(state: OverallState):
    return [Send("generate_joke", {"subject": s}) for s in state['subjects']]

graph.add_conditional_edges("node_a", continue_to_jokes)

Checkpointer

LangGraph has a built-in persistence layer, implemented through [checkpointers][basecheckpointsaver]. When you use a checkpointer with a graph, you can interact with the state of that graph. When you use a checkpointer with a graph, you can interact with and manage the graph's state. The checkpointer saves a checkpoint of the graph state at every super-step, enabling several powerful capabilities:

First, checkpointers facilitate human-in-the-loop workflows workflows by allowing humans to inspect, interrupt, and approve steps.Checkpointers are needed for these workflows as the human has to be able to view the state of a graph at any point in time, and the graph has to be to resume execution after the human has made any updates to the state.

Second, it allows for "memory" between interactions. You can use checkpointers to create threads and save the state of a thread after a graph executes. In the case of repeated human interactions (like conversations) any follow up messages can be sent to that checkpoint, which will retain its memory of previous ones.

See this guide for how to add a checkpointer to your graph.

Threads

Threads enable the checkpointing of multiple different runs, making them essential for multi-tenant chat applications and other scenarios where maintaining separate states is necessary. A thread is a unique ID assigned to a series of checkpoints saved by a checkpointer. When using a checkpointer, you must specify a thread_id or thread_ts when running the graph.

thread_id is simply the ID of a thread. This is always required

thread_ts can optionally be passed. This identifier refers to a specific checkpoint within a thread. This can be used to kick of a run of a graph from some point halfway through a thread.

You must pass these when invoking the graph as part of the configurable part of the config.

config = {"configurable": {"thread_id": "a"}}
graph.invoke(inputs, config=config)

See this guide for how to use threads.

Checkpointer state

When interacting with the checkpointer state, you must specify a thread identifier.Each checkpoint saved by the checkpointer has two properties:

• values: This is the value of the state at this point in time.
• next: This is a tuple of the nodes to execute next in the graph.

Get state

You can get the state of a checkpointer by calling graph.get_state(config). The config should contain thread_id, and the state will be fetched for that thread.

Get state history

You can also call graph.get_state_history(config) to get a list of the history of the graph. The config should contain thread_id, and the state history will be fetched for that thread.

Update state

You can also interact with the state directly and update it. This takes three different components:

• config
• values
• as_node

config

The config should contain thread_id specifying which thread to update.

values

These are the values that will be used to update the state. Note that this update is treated exactly as any update from a node is treated. This means that these values will be passed to the reducer functions that are part of the state. So this does NOT automatically overwrite the state. Let's walk through an example.

Let's assume you have defined the state of your graph as:

from typing import TypedDict, Annotated
from operator import add

class State(TypedDict):
    foo: int
    bar: Annotated[list[str], add]

Let's now assume the current state of the graph is

{"foo": 1, "bar": ["a"]}

If you update the state as below:

graph.update_state(config, {"foo": 2, "bar": ["b"]})

Then the new state of the graph will be:

{"foo": 2, "bar": ["a", "b"]}

The foo key is completely changed (because there is no reducer specified for that key, so it overwrites it). However, there is a reducer specified for the bar key, and so it appends "b" to the state of bar.

as_node

The final thing you specify when calling update_state is as_node. This update will be applied as if it came from node as_node. If as_node is not provided, it will be set to the last node that updated the state, if not ambiguous.

The reason this matters is that the next steps in the graph to execute depend on the last node to have given an update, so this can be used to control which node executes next.

Graph Migrations

LangGraph can easily handle migrations of graph definitions (nodes, edges, and state) even when using a checkpointer to track state.

• For threads at the end of the graph (i.e. not interrupted) you can change the entire topology of the graph (i.e. all nodes and edges, remove, add, rename, etc)
• For threads currently interrupted, we support all topology changes other than renaming / removing nodes (as that thread could now be about to enter a node that no longer exists) -- if this is a blocker please reach out and we can prioritize a solution.
• For modifying state, we have full backwards and forwards compatibility for adding and removing keys
• State keys that are renamed lose their saved state in existing threads
• State keys whose types change in incompatible ways could currently cause issues in threads with state from before the change -- if this is a blocker please reach out and we can prioritize a solution.

Configuration

When creating a graph, you can also mark that certain parts of the graph are configurable. This is commonly done to enable easily switching between models or system prompts. This allows you to create a single "cognitive architecture" (the graph) but have multiple different instance of it.

You can optionally specify a config_schema when creating a graph.

class ConfigSchema(TypedDict):
    llm: str

graph = StateGraph(State, config_schema=ConfigSchema)

You can then pass this configuration into the graph using the configurable config field.

config = {"configurable": {"llm": "anthropic"}}

graph.invoke(inputs, config=config)

You can then access and use this configuration inside a node:

def node_a(state, config):
    llm_type = config.get("configurable", {}).get("llm", "openai")
    llm = get_llm(llm_type)
    ...

See this guide for a full breakdown on configuration

Breakpoints

It can often be useful to set breakpoints before or after certain nodes execute. This can be used to wait for human approval before continuing. These can be set when you "compile" a graph. You can set breakpoints either before a node executes (using interrupt_before) or after a node executes (using interrupt_after.)

You MUST use a checkpoiner when using breakpoints. This is because your graph needs to be able to resume execution.

In order to resume execution, you can just invoke your graph with None as the input.

# Initial run of graph
graph.invoke(inputs, config=config)

# Let's assume it hit a breakpoint somewhere, you can then resume by passing in None
graph.invoke(None, config=config)

See this guide for a full walkthrough of how to add breakpoints.

Visualization

It's often nice to be able to visualize graphs, especially as they get more complex. LangGraph comes with several built-in ways to visualize graphs. See this how-to guide for more info.

Streaming

LangGraph is built with first class support for streaming. There are several different streaming modes that LangGraph supports:

• "values": This streams the full value of the state after each step of the graph.
• "updates: This streams the updates to the state after each step of the graph. If multiple updates are made in the same step (e.g. multiple nodes are run) then those updates are streamed separately.
• "debug": This streams as much information as possible throughout the execution of the graph.

In addition, you can use the astream_events method to stream back events that happen inside nodes. This is useful for streaming tokens of LLM calls.

!!! warning "ASYNC IN PYTHON<=3.10" You may fail to see events being emitted from inside a node when using .astream_events in Python <= 3.10. If you're using a Langchain RunnableLambda, a RunnableGenerator, or Tool asynchronously inside your node, you will have to propagate callbacks to these objects manually. This is because LangChain cannot automatically propagate callbacks to child objects in this case. Please see examples here and here.