diff --git a/examples/rag/langgraph_agentic_rag.ipynb b/examples/rag/langgraph_agentic_rag.ipynb index 486c6caed..73d02f14a 100644 --- a/examples/rag/langgraph_agentic_rag.ipynb +++ b/examples/rag/langgraph_agentic_rag.ipynb @@ -20,7 +20,10 @@ "id": "969fb438", "metadata": {}, "outputs": [], - "source": ["%%capture --no-stderr\n%pip install -U --quiet langchain-community tiktoken langchain-openai langchainhub chromadb langchain langgraph langchain-text-splitters"] + "source": [ + "%%capture --no-stderr\n", + "%pip install -U --quiet langchain-community tiktoken langchain-openai langchainhub chromadb langchain langgraph langchain-text-splitters" + ] }, { "cell_type": "code", @@ -28,7 +31,22 @@ "id": "e4958a8c", "metadata": {}, "outputs": [], - "source": ["import getpass\nimport os\n\n\ndef _set_env(key: str):\n if key not in os.environ:\n os.environ[key] = getpass.getpass(f\"{key}:\")\n\n\n_set_env(\"OPENAI_API_KEY\")\n\n# (Optional) For tracing\nos.environ[\"LANGCHAIN_TRACING_V2\"] = \"true\"\n_set_env(\"LANGCHAIN_API_KEY\")"] + "source": [ + "import getpass\n", + "import os\n", + "\n", + "\n", + "def _set_env(key: str):\n", + " if key not in os.environ:\n", + " os.environ[key] = getpass.getpass(f\"{key}:\")\n", + "\n", + "\n", + "_set_env(\"OPENAI_API_KEY\")\n", + "\n", + "# (Optional) For tracing\n", + "os.environ[\"LANGCHAIN_TRACING_V2\"] = \"true\"\n", + "_set_env(\"LANGCHAIN_API_KEY\")" + ] }, { "cell_type": "markdown", @@ -46,7 +64,34 @@ "id": "e50c9efe-4abe-42fa-b35a-05eeeede9ec6", "metadata": {}, "outputs": [], - "source": ["from langchain_community.document_loaders import WebBaseLoader\nfrom langchain_community.vectorstores import Chroma\nfrom langchain_openai import OpenAIEmbeddings\nfrom langchain_text_splitters import RecursiveCharacterTextSplitter\n\nurls = [\n \"https://lilianweng.github.io/posts/2023-06-23-agent/\",\n \"https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/\",\n \"https://lilianweng.github.io/posts/2023-10-25-adv-attack-llm/\",\n]\n\ndocs = [WebBaseLoader(url).load() for url in urls]\ndocs_list = [item for sublist in docs for item in sublist]\n\ntext_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(\n chunk_size=100, chunk_overlap=50\n)\ndoc_splits = text_splitter.split_documents(docs_list)\n\n# Add to vectorDB\nvectorstore = Chroma.from_documents(\n documents=doc_splits,\n collection_name=\"rag-chroma\",\n embedding=OpenAIEmbeddings(),\n)\nretriever = vectorstore.as_retriever()"] + "source": [ + "from langchain_community.document_loaders import WebBaseLoader\n", + "from langchain_community.vectorstores import Chroma\n", + "from langchain_openai import OpenAIEmbeddings\n", + "from langchain_text_splitters import RecursiveCharacterTextSplitter\n", + "\n", + "urls = [\n", + " \"https://lilianweng.github.io/posts/2023-06-23-agent/\",\n", + " \"https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/\",\n", + " \"https://lilianweng.github.io/posts/2023-10-25-adv-attack-llm/\",\n", + "]\n", + "\n", + "docs = [WebBaseLoader(url).load() for url in urls]\n", + "docs_list = [item for sublist in docs for item in sublist]\n", + "\n", + "text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(\n", + " chunk_size=100, chunk_overlap=50\n", + ")\n", + "doc_splits = text_splitter.split_documents(docs_list)\n", + "\n", + "# Add to vectorDB\n", + "vectorstore = Chroma.from_documents(\n", + " documents=doc_splits,\n", + " collection_name=\"rag-chroma\",\n", + " embedding=OpenAIEmbeddings(),\n", + ")\n", + "retriever = vectorstore.as_retriever()" + ] }, { "cell_type": "markdown", @@ -62,7 +107,17 @@ "id": "0b97bdd8-d7e3-444d-ac96-5ef4725f9048", "metadata": {}, "outputs": [], - "source": ["from langchain.tools.retriever import create_retriever_tool\n\nretriever_tool = create_retriever_tool(\n retriever,\n \"retrieve_blog_posts\",\n \"Search and return information about Lilian Weng blog posts on LLM agents, prompt engineering, and adversarial attacks on LLMs.\",\n)\n\ntools = [retriever_tool]"] + "source": [ + "from langchain.tools.retriever import create_retriever_tool\n", + "\n", + "retriever_tool = create_retriever_tool(\n", + " retriever,\n", + " \"retrieve_blog_posts\",\n", + " \"Search and return information about Lilian Weng blog posts on LLM agents, prompt engineering, and adversarial attacks on LLMs.\",\n", + ")\n", + "\n", + "tools = [retriever_tool]" + ] }, { "cell_type": "markdown", @@ -86,7 +141,19 @@ "id": "0e378706-47d5-425a-8ba0-57b9acffbd0c", "metadata": {}, "outputs": [], - "source": ["from typing import Annotated, Sequence, TypedDict\n\nfrom langchain_core.messages import BaseMessage\n\nfrom langgraph.graph.message import add_messages\n\n\nclass AgentState(TypedDict):\n # The add_messages function defines how an update should be processed\n # Default is to replace. add_messages says \"append\"\n messages: Annotated[Sequence[BaseMessage], add_messages]"] + "source": [ + "from typing import Annotated, Sequence, TypedDict\n", + "\n", + "from langchain_core.messages import BaseMessage\n", + "\n", + "from langgraph.graph.message import add_messages\n", + "\n", + "\n", + "class AgentState(TypedDict):\n", + " # The add_messages function defines how an update should be processed\n", + " # Default is to replace. add_messages says \"append\"\n", + " messages: Annotated[Sequence[BaseMessage], add_messages]" + ] }, { "attachments": { @@ -129,7 +196,173 @@ ] } ], - "source": ["from typing import Annotated, Literal, Sequence, TypedDict\n\nfrom langchain import hub\nfrom langchain_core.messages import BaseMessage, HumanMessage\nfrom langchain_core.output_parsers import StrOutputParser\nfrom langchain_core.prompts import PromptTemplate\nfrom langchain_core.pydantic_v1 import BaseModel, Field\nfrom langchain_openai import ChatOpenAI\n\nfrom langgraph.prebuilt import tools_condition\n\n### Edges\n\n\ndef grade_documents(state) -> Literal[\"generate\", \"rewrite\"]:\n \"\"\"\n Determines whether the retrieved documents are relevant to the question.\n\n Args:\n state (messages): The current state\n\n Returns:\n str: A decision for whether the documents are relevant or not\n \"\"\"\n\n print(\"---CHECK RELEVANCE---\")\n\n # Data model\n class grade(BaseModel):\n \"\"\"Binary score for relevance check.\"\"\"\n\n binary_score: str = Field(description=\"Relevance score 'yes' or 'no'\")\n\n # LLM\n model = ChatOpenAI(temperature=0, model=\"gpt-4-0125-preview\", streaming=True)\n\n # LLM with tool and validation\n llm_with_tool = model.with_structured_output(grade)\n\n # Prompt\n prompt = PromptTemplate(\n template=\"\"\"You are a grader assessing relevance of a retrieved document to a user question. \\n \n Here is the retrieved document: \\n\\n {context} \\n\\n\n Here is the user question: {question} \\n\n If the document contains keyword(s) or semantic meaning related to the user question, grade it as relevant. \\n\n Give a binary score 'yes' or 'no' score to indicate whether the document is relevant to the question.\"\"\",\n input_variables=[\"context\", \"question\"],\n )\n\n # Chain\n chain = prompt | llm_with_tool\n\n messages = state[\"messages\"]\n last_message = messages[-1]\n\n question = messages[0].content\n docs = last_message.content\n\n scored_result = chain.invoke({\"question\": question, \"context\": docs})\n\n score = scored_result.binary_score\n\n if score == \"yes\":\n print(\"---DECISION: DOCS RELEVANT---\")\n return \"generate\"\n\n else:\n print(\"---DECISION: DOCS NOT RELEVANT---\")\n print(score)\n return \"rewrite\"\n\n\n### Nodes\n\n\ndef agent(state):\n \"\"\"\n Invokes the agent model to generate a response based on the current state. Given\n the question, it will decide to retrieve using the retriever tool, or simply end.\n\n Args:\n state (messages): The current state\n\n Returns:\n dict: The updated state with the agent response appended to messages\n \"\"\"\n print(\"---CALL AGENT---\")\n messages = state[\"messages\"]\n model = ChatOpenAI(temperature=0, streaming=True, model=\"gpt-4-turbo\")\n model = model.bind_tools(tools)\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\ndef rewrite(state):\n \"\"\"\n Transform the query to produce a better question.\n\n Args:\n state (messages): The current state\n\n Returns:\n dict: The updated state with re-phrased question\n \"\"\"\n\n print(\"---TRANSFORM QUERY---\")\n messages = state[\"messages\"]\n question = messages[0].content\n\n msg = [\n HumanMessage(\n content=f\"\"\" \\n \n Look at the input and try to reason about the underlying semantic intent / meaning. \\n \n Here is the initial question:\n \\n ------- \\n\n {question} \n \\n ------- \\n\n Formulate an improved question: \"\"\",\n )\n ]\n\n # Grader\n model = ChatOpenAI(temperature=0, model=\"gpt-4-0125-preview\", streaming=True)\n response = model.invoke(msg)\n return {\"messages\": [response]}\n\n\ndef generate(state):\n \"\"\"\n Generate answer\n\n Args:\n state (messages): The current state\n\n Returns:\n dict: The updated state with re-phrased question\n \"\"\"\n print(\"---GENERATE---\")\n messages = state[\"messages\"]\n question = messages[0].content\n last_message = messages[-1]\n\n question = messages[0].content\n docs = last_message.content\n\n # Prompt\n prompt = hub.pull(\"rlm/rag-prompt\")\n\n # LLM\n llm = ChatOpenAI(model_name=\"gpt-3.5-turbo\", temperature=0, streaming=True)\n\n # Post-processing\n def format_docs(docs):\n return \"\\n\\n\".join(doc.page_content for doc in docs)\n\n # Chain\n rag_chain = prompt | llm | StrOutputParser()\n\n # Run\n response = rag_chain.invoke({\"context\": docs, \"question\": question})\n return {\"messages\": [response]}\n\n\nprint(\"*\" * 20 + \"Prompt[rlm/rag-prompt]\" + \"*\" * 20)\nprompt = hub.pull(\"rlm/rag-prompt\").pretty_print() # Show what the prompt looks like"] + "source": [ + "from typing import Annotated, Literal, Sequence, TypedDict\n", + "\n", + "from langchain import hub\n", + "from langchain_core.messages import BaseMessage, HumanMessage\n", + "from langchain_core.output_parsers import StrOutputParser\n", + "from langchain_core.prompts import PromptTemplate\n", + "from langchain_core.pydantic_v1 import BaseModel, Field\n", + "from langchain_openai import ChatOpenAI\n", + "\n", + "from langgraph.prebuilt import tools_condition\n", + "\n", + "### Edges\n", + "\n", + "\n", + "def grade_documents(state) -> Literal[\"generate\", \"rewrite\"]:\n", + " \"\"\"\n", + " Determines whether the retrieved documents are relevant to the question.\n", + "\n", + " Args:\n", + " state (messages): The current state\n", + "\n", + " Returns:\n", + " str: A decision for whether the documents are relevant or not\n", + " \"\"\"\n", + "\n", + " print(\"---CHECK RELEVANCE---\")\n", + "\n", + " # Data model\n", + " class grade(BaseModel):\n", + " \"\"\"Binary score for relevance check.\"\"\"\n", + "\n", + " binary_score: str = Field(description=\"Relevance score 'yes' or 'no'\")\n", + "\n", + " # LLM\n", + " model = ChatOpenAI(temperature=0, model=\"gpt-4-0125-preview\", streaming=True)\n", + "\n", + " # LLM with tool and validation\n", + " llm_with_tool = model.with_structured_output(grade)\n", + "\n", + " # Prompt\n", + " prompt = PromptTemplate(\n", + " template=\"\"\"You are a grader assessing relevance of a retrieved document to a user question. \\n \n", + " Here is the retrieved document: \\n\\n {context} \\n\\n\n", + " Here is the user question: {question} \\n\n", + " If the document contains keyword(s) or semantic meaning related to the user question, grade it as relevant. \\n\n", + " Give a binary score 'yes' or 'no' score to indicate whether the document is relevant to the question.\"\"\",\n", + " input_variables=[\"context\", \"question\"],\n", + " )\n", + "\n", + " # Chain\n", + " chain = prompt | llm_with_tool\n", + "\n", + " messages = state[\"messages\"]\n", + " last_message = messages[-1]\n", + "\n", + " question = messages[0].content\n", + " docs = last_message.content\n", + "\n", + " scored_result = chain.invoke({\"question\": question, \"context\": docs})\n", + "\n", + " score = scored_result.binary_score\n", + "\n", + " if score == \"yes\":\n", + " print(\"---DECISION: DOCS RELEVANT---\")\n", + " return \"generate\"\n", + "\n", + " else:\n", + " print(\"---DECISION: DOCS NOT RELEVANT---\")\n", + " print(score)\n", + " return \"rewrite\"\n", + "\n", + "\n", + "### Nodes\n", + "\n", + "\n", + "def agent(state):\n", + " \"\"\"\n", + " Invokes the agent model to generate a response based on the current state. Given\n", + " the question, it will decide to retrieve using the retriever tool, or simply end.\n", + "\n", + " Args:\n", + " state (messages): The current state\n", + "\n", + " Returns:\n", + " dict: The updated state with the agent response appended to messages\n", + " \"\"\"\n", + " print(\"---CALL AGENT---\")\n", + " messages = state[\"messages\"]\n", + " model = ChatOpenAI(temperature=0, streaming=True, model=\"gpt-4-turbo\")\n", + " model = model.bind_tools(tools)\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", + "def rewrite(state):\n", + " \"\"\"\n", + " Transform the query to produce a better question.\n", + "\n", + " Args:\n", + " state (messages): The current state\n", + "\n", + " Returns:\n", + " dict: The updated state with re-phrased question\n", + " \"\"\"\n", + "\n", + " print(\"---TRANSFORM QUERY---\")\n", + " messages = state[\"messages\"]\n", + " question = messages[0].content\n", + "\n", + " msg = [\n", + " HumanMessage(\n", + " content=f\"\"\" \\n \n", + " Look at the input and try to reason about the underlying semantic intent / meaning. \\n \n", + " Here is the initial question:\n", + " \\n ------- \\n\n", + " {question} \n", + " \\n ------- \\n\n", + " Formulate an improved question: \"\"\",\n", + " )\n", + " ]\n", + "\n", + " # Grader\n", + " model = ChatOpenAI(temperature=0, model=\"gpt-4-0125-preview\", streaming=True)\n", + " response = model.invoke(msg)\n", + " return {\"messages\": [response]}\n", + "\n", + "\n", + "def generate(state):\n", + " \"\"\"\n", + " Generate answer\n", + "\n", + " Args:\n", + " state (messages): The current state\n", + "\n", + " Returns:\n", + " dict: The updated state with re-phrased question\n", + " \"\"\"\n", + " print(\"---GENERATE---\")\n", + " messages = state[\"messages\"]\n", + " question = messages[0].content\n", + " last_message = messages[-1]\n", + "\n", + " docs = last_message.content\n", + "\n", + " # Prompt\n", + " prompt = hub.pull(\"rlm/rag-prompt\")\n", + "\n", + " # LLM\n", + " llm = ChatOpenAI(model_name=\"gpt-3.5-turbo\", temperature=0, streaming=True)\n", + "\n", + " # Post-processing\n", + " def format_docs(docs):\n", + " return \"\\n\\n\".join(doc.page_content for doc in docs)\n", + "\n", + " # Chain\n", + " rag_chain = prompt | llm | StrOutputParser()\n", + "\n", + " # Run\n", + " response = rag_chain.invoke({\"context\": docs, \"question\": question})\n", + " return {\"messages\": [response]}\n", + "\n", + "\n", + "print(\"*\" * 20 + \"Prompt[rlm/rag-prompt]\" + \"*\" * 20)\n", + "prompt = hub.pull(\"rlm/rag-prompt\").pretty_print() # Show what the prompt looks like" + ] }, { "cell_type": "markdown", @@ -150,7 +383,48 @@ "id": "8718a37f-83c2-4f16-9850-e61e0f49c3d4", "metadata": {}, "outputs": [], - "source": ["from langgraph.graph import END, StateGraph, START\nfrom langgraph.prebuilt import ToolNode\n\n# Define a new graph\nworkflow = StateGraph(AgentState)\n\n# Define the nodes we will cycle between\nworkflow.add_node(\"agent\", agent) # agent\nretrieve = ToolNode([retriever_tool])\nworkflow.add_node(\"retrieve\", retrieve) # retrieval\nworkflow.add_node(\"rewrite\", rewrite) # Re-writing the question\nworkflow.add_node(\n \"generate\", generate\n) # Generating a response after we know the documents are relevant\n# Call agent node to decide to retrieve or not\nworkflow.add_edge(START, \"agent\")\n\n# Decide whether to retrieve\nworkflow.add_conditional_edges(\n \"agent\",\n # Assess agent decision\n tools_condition,\n {\n # Translate the condition outputs to nodes in our graph\n \"tools\": \"retrieve\",\n END: END,\n },\n)\n\n# Edges taken after the `action` node is called.\nworkflow.add_conditional_edges(\n \"retrieve\",\n # Assess agent decision\n grade_documents,\n)\nworkflow.add_edge(\"generate\", END)\nworkflow.add_edge(\"rewrite\", \"agent\")\n\n# Compile\ngraph = workflow.compile()"] + "source": [ + "from langgraph.graph import END, StateGraph, START\n", + "from langgraph.prebuilt import ToolNode\n", + "\n", + "# Define a new graph\n", + "workflow = StateGraph(AgentState)\n", + "\n", + "# Define the nodes we will cycle between\n", + "workflow.add_node(\"agent\", agent) # agent\n", + "retrieve = ToolNode([retriever_tool])\n", + "workflow.add_node(\"retrieve\", retrieve) # retrieval\n", + "workflow.add_node(\"rewrite\", rewrite) # Re-writing the question\n", + "workflow.add_node(\n", + " \"generate\", generate\n", + ") # Generating a response after we know the documents are relevant\n", + "# Call agent node to decide to retrieve or not\n", + "workflow.add_edge(START, \"agent\")\n", + "\n", + "# Decide whether to retrieve\n", + "workflow.add_conditional_edges(\n", + " \"agent\",\n", + " # Assess agent decision\n", + " tools_condition,\n", + " {\n", + " # Translate the condition outputs to nodes in our graph\n", + " \"tools\": \"retrieve\",\n", + " END: END,\n", + " },\n", + ")\n", + "\n", + "# Edges taken after the `action` node is called.\n", + "workflow.add_conditional_edges(\n", + " \"retrieve\",\n", + " # Assess agent decision\n", + " grade_documents,\n", + ")\n", + "workflow.add_edge(\"generate\", END)\n", + "workflow.add_edge(\"rewrite\", \"agent\")\n", + "\n", + "# Compile\n", + "graph = workflow.compile()" + ] }, { "cell_type": "code", @@ -169,7 +443,15 @@ "output_type": "display_data" } ], - "source": ["from IPython.display import Image, display\n\ntry:\n display(Image(graph.get_graph(xray=True).draw_mermaid_png()))\nexcept Exception:\n # This requires some extra dependencies and is optional\n pass"] + "source": [ + "from IPython.display import Image, display\n", + "\n", + "try:\n", + " display(Image(graph.get_graph(xray=True).draw_mermaid_png()))\n", + "except Exception:\n", + " # This requires some extra dependencies and is optional\n", + " pass" + ] }, { "cell_type": "code", @@ -203,7 +485,21 @@ ] } ], - "source": ["import pprint\n\ninputs = {\n \"messages\": [\n (\"user\", \"What does Lilian Weng say about the types of agent memory?\"),\n ]\n}\nfor output in graph.stream(inputs):\n for key, value in output.items():\n pprint.pprint(f\"Output from node '{key}':\")\n pprint.pprint(\"---\")\n pprint.pprint(value, indent=2, width=80, depth=None)\n pprint.pprint(\"\\n---\\n\")"] + "source": [ + "import pprint\n", + "\n", + "inputs = {\n", + " \"messages\": [\n", + " (\"user\", \"What does Lilian Weng say about the types of agent memory?\"),\n", + " ]\n", + "}\n", + "for output in graph.stream(inputs):\n", + " for key, value in output.items():\n", + " pprint.pprint(f\"Output from node '{key}':\")\n", + " pprint.pprint(\"---\")\n", + " pprint.pprint(value, indent=2, width=80, depth=None)\n", + " pprint.pprint(\"\\n---\\n\")" + ] }, { "cell_type": "code", @@ -211,7 +507,7 @@ "id": "189333cc-5d34-4869-9f9b-741210e1096f", "metadata": {}, "outputs": [], - "source": [""] + "source": [] } ], "metadata": {