代理式 RAG¶

在本教程中，我们将构建一个检索代理。当您希望 LLM 决定是从向量存储中检索上下文还是直接响应用户时，检索代理非常有用。

在本教程结束时，我们将完成以下工作：

1. 获取并预处理将用于检索的文档。
2. 对这些文档进行索引以进行语义搜索，并为代理创建一个检索器工具。
3. 构建一个代理式 RAG 系统，该系统可以决定何时使用检索器工具。

设置¶

让我们下载所需的包并设置我们的 API 密钥

%%capture --no-stderr
%pip install -U --quiet langgraph "langchain[openai]" langchain-community langchain-text-splitters

import getpass
import os


def _set_env(key: str):
    if key not in os.environ:
        os.environ[key] = getpass.getpass(f"{key}:")


_set_env("OPENAI_API_KEY")

1. 预处理文档¶

1. 获取要在我们的 RAG 系统中使用的文档。我们将使用Lilian Weng 优秀博客中最近的三页。我们将首先使用WebBaseLoader工具获取页面的内容

   from langchain_community.document_loaders import WebBaseLoader
   
   urls = [
       "https://lilianweng.github.io/posts/2024-11-28-reward-hacking/",
       "https://lilianweng.github.io/posts/2024-07-07-hallucination/",
       "https://lilianweng.github.io/posts/2024-04-12-diffusion-video/",
   ]
   
   docs = [WebBaseLoader(url).load() for url in urls]
   

   docs[0][0].page_content.strip()[:1000]
   

2. 将获取的文档分割成更小的块，以便索引到我们的向量存储中

   from langchain_text_splitters import RecursiveCharacterTextSplitter
   
   docs_list = [item for sublist in docs for item in sublist]
   
   text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(
       chunk_size=100, chunk_overlap=50
   )
   doc_splits = text_splitter.split_documents(docs_list)
   

   doc_splits[0].page_content.strip()
   

2. 创建检索器工具¶

现在我们已经分割了文档，我们可以将它们索引到我们将用于语义搜索的向量存储中。

1. 使用内存向量存储和 OpenAI 嵌入

   from langchain_core.vectorstores import InMemoryVectorStore
   from langchain_openai import OpenAIEmbeddings
   
   vectorstore = InMemoryVectorStore.from_documents(
       documents=doc_splits, embedding=OpenAIEmbeddings()
   )
   retriever = vectorstore.as_retriever()
   

2. 使用 LangChain 预构建的 create_retriever_tool 创建一个检索器工具

   from langchain.tools.retriever import create_retriever_tool
   
   retriever_tool = create_retriever_tool(
       retriever,
       "retrieve_blog_posts",
       "Search and return information about Lilian Weng blog posts.",
   )
   

3. 测试工具

   retriever_tool.invoke({"query": "types of reward hacking"})
   

3. 生成查询¶

现在我们将开始为我们的代理式 RAG 图构建组件（节点和边）。请注意，这些组件将作用于 MessagesState — 包含一个带有聊天消息列表的messages键的图状态。

1. 构建一个 generate_query_or_respond 节点。它将调用 LLM 根据当前的图状态（消息列表）生成响应。给定输入消息，它将决定使用检索器工具进行检索，或直接响应用户。请注意，我们通过.bind_tools让聊天模型访问我们之前创建的retriever_tool

   from langgraph.graph import MessagesState
   from langchain.chat_models import init_chat_model
   
   response_model = init_chat_model("openai:gpt-4.1", temperature=0)
   
   
   def generate_query_or_respond(state: MessagesState):
       """Call the model to generate a response based on the current state. Given
       the question, it will decide to retrieve using the retriever tool, or simply respond to the user.
       """
       response = (
           response_model
           .bind_tools([retriever_tool]).invoke(state["messages"])
       )
       return {"messages": [response]}
   

2. 在随机输入上尝试

   input = {"messages": [{"role": "user", "content": "hello!"}]}
   generate_query_or_respond(input)["messages"][-1].pretty_print()
   

   输出

   ================================== Ai Message ==================================
   
   Hello! How can I help you today?
   

3. 提出一个需要语义搜索的问题

   input = {
       "messages": [
           {
               "role": "user",
               "content": "What does Lilian Weng say about types of reward hacking?",
           }
       ]
   }
   generate_query_or_respond(input)["messages"][-1].pretty_print()
   

   输出

   ================================== Ai Message ==================================
   Tool Calls:
   retrieve_blog_posts (call_tYQxgfIlnQUDMdtAhdbXNwIM)
   Call ID: call_tYQxgfIlnQUDMdtAhdbXNwIM
   Args:
       query: types of reward hacking
   

4. 评估文档¶

1. 添加一个条件边 — grade_documents — 以确定检索到的文档是否与问题相关。我们将使用具有结构化输出模式GradeDocuments的模型进行文档评估。grade_documents函数将根据评估决定（generate_answer或rewrite_question）返回要前往的节点名称

   from pydantic import BaseModel, Field
   from typing import Literal
   
   GRADE_PROMPT = (
       "You are a grader assessing relevance of a retrieved document to a user question. \n "
       "Here is the retrieved document: \n\n {context} \n\n"
       "Here is the user question: {question} \n"
       "If the document contains keyword(s) or semantic meaning related to the user question, grade it as relevant. \n"
       "Give a binary score 'yes' or 'no' score to indicate whether the document is relevant to the question."
   )
   
   
   class GradeDocuments(BaseModel):
       """Grade documents using a binary score for relevance check."""
   
       binary_score: str = Field(
           description="Relevance score: 'yes' if relevant, or 'no' if not relevant"
       )
   
   
   grader_model = init_chat_model("openai:gpt-4.1", temperature=0)
   
   
   def grade_documents(
       state: MessagesState,
   ) -> Literal["generate_answer", "rewrite_question"]:
       """Determine whether the retrieved documents are relevant to the question."""
       question = state["messages"][0].content
       context = state["messages"][-1].content
   
       prompt = GRADE_PROMPT.format(question=question, context=context)
       response = (
           grader_model
           .with_structured_output(GradeDocuments).invoke(
               [{"role": "user", "content": prompt}]
           )
       )
       score = response.binary_score
   
       if score == "yes":
           return "generate_answer"
       else:
           return "rewrite_question"
   

2. 在工具响应中包含不相关文档的情况下运行此操作

   from langchain_core.messages import convert_to_messages
   
   input = {
       "messages": convert_to_messages(
           [
               {
                   "role": "user",
                   "content": "What does Lilian Weng say about types of reward hacking?",
               },
               {
                   "role": "assistant",
                   "content": "",
                   "tool_calls": [
                       {
                           "id": "1",
                           "name": "retrieve_blog_posts",
                           "args": {"query": "types of reward hacking"},
                       }
                   ],
               },
               {"role": "tool", "content": "meow", "tool_call_id": "1"},
           ]
       )
   }
   grade_documents(input)
   

3. 确认相关文档已分类

   input = {
       "messages": convert_to_messages(
           [
               {
                   "role": "user",
                   "content": "What does Lilian Weng say about types of reward hacking?",
               },
               {
                   "role": "assistant",
                   "content": "",
                   "tool_calls": [
                       {
                           "id": "1",
                           "name": "retrieve_blog_posts",
                           "args": {"query": "types of reward hacking"},
                       }
                   ],
               },
               {
                   "role": "tool",
                   "content": "reward hacking can be categorized into two types: environment or goal misspecification, and reward tampering",
                   "tool_call_id": "1",
               },
           ]
       )
   }
   grade_documents(input)
   

5. 重写问题¶

1. 构建 rewrite_question 节点。检索器工具可能会返回潜在不相关的文档，这表明需要改进原始用户问题。为此，我们将调用 rewrite_question 节点

   REWRITE_PROMPT = (
       "Look at the input and try to reason about the underlying semantic intent / meaning.\n"
       "Here is the initial question:"
       "\n ------- \n"
       "{question}"
       "\n ------- \n"
       "Formulate an improved question:"
   )
   
   
   def rewrite_question(state: MessagesState):
       """Rewrite the original user question."""
       messages = state["messages"]
       question = messages[0].content
       prompt = REWRITE_PROMPT.format(question=question)
       response = response_model.invoke([{"role": "user", "content": prompt}])
       return {"messages": [{"role": "user", "content": response.content}]}
   

2. 尝试一下

   input = {
       "messages": convert_to_messages(
           [
               {
                   "role": "user",
                   "content": "What does Lilian Weng say about types of reward hacking?",
               },
               {
                   "role": "assistant",
                   "content": "",
                   "tool_calls": [
                       {
                           "id": "1",
                           "name": "retrieve_blog_posts",
                           "args": {"query": "types of reward hacking"},
                       }
                   ],
               },
               {"role": "tool", "content": "meow", "tool_call_id": "1"},
           ]
       )
   }
   
   response = rewrite_question(input)
   print(response["messages"][-1]["content"])
   

   输出

   What are the different types of reward hacking described by Lilian Weng, and how does she explain them?
   

6. 生成答案¶

1. 构建 generate_answer 节点：如果我们通过了评估器检查，我们可以根据原始问题和检索到的上下文生成最终答案

   GENERATE_PROMPT = (
       "You are an assistant for question-answering tasks. "
       "Use the following pieces of retrieved context to answer the question. "
       "If you don't know the answer, just say that you don't know. "
       "Use three sentences maximum and keep the answer concise.\n"
       "Question: {question} \n"
       "Context: {context}"
   )
   
   
   def generate_answer(state: MessagesState):
       """Generate an answer."""
       question = state["messages"][0].content
       context = state["messages"][-1].content
       prompt = GENERATE_PROMPT.format(question=question, context=context)
       response = response_model.invoke([{"role": "user", "content": prompt}])
       return {"messages": [response]}
   

2. 尝试一下

   input = {
       "messages": convert_to_messages(
           [
               {
                   "role": "user",
                   "content": "What does Lilian Weng say about types of reward hacking?",
               },
               {
                   "role": "assistant",
                   "content": "",
                   "tool_calls": [
                       {
                           "id": "1",
                           "name": "retrieve_blog_posts",
                           "args": {"query": "types of reward hacking"},
                       }
                   ],
               },
               {
                   "role": "tool",
                   "content": "reward hacking can be categorized into two types: environment or goal misspecification, and reward tampering",
                   "tool_call_id": "1",
               },
           ]
       )
   }
   
   response = generate_answer(input)
   response["messages"][-1].pretty_print()
   

   输出

   ================================== Ai Message ==================================
   
   Lilian Weng categorizes reward hacking into two types: environment or goal misspecification, and reward tampering. She considers reward hacking as a broad concept that includes both of these categories. Reward hacking occurs when an agent exploits flaws or ambiguities in the reward function to achieve high rewards without performing the intended behaviors.
   

7. 组装图¶

• 从 generate_query_or_respond 开始，并确定是否需要调用 retriever_tool
• 使用 tools_condition 路由到下一步
  • 如果 generate_query_or_respond 返回了 tool_calls，则调用 retriever_tool 检索上下文
  • 否则，直接响应用户
• 评估检索到的文档内容与问题的相关性（grade_documents），并路由到下一步
  • 如果不相关，则使用 rewrite_question 重写问题，然后再次调用 generate_query_or_respond
  • 如果相关，则继续到 generate_answer 并使用包含检索到的文档上下文的 ToolMessage 生成最终响应

^{API 参考：StateGraph | START | END | ToolNode | tools_condition}

from langgraph.graph import StateGraph, START, END
from langgraph.prebuilt import ToolNode
from langgraph.prebuilt import tools_condition

workflow = StateGraph(MessagesState)

# Define the nodes we will cycle between
workflow.add_node(generate_query_or_respond)
workflow.add_node("retrieve", ToolNode([retriever_tool]))
workflow.add_node(rewrite_question)
workflow.add_node(generate_answer)

workflow.add_edge(START, "generate_query_or_respond")

# Decide whether to retrieve
workflow.add_conditional_edges(
    "generate_query_or_respond",
    # Assess LLM decision (call `retriever_tool` tool or respond to the user)
    tools_condition,
    {
        # Translate the condition outputs to nodes in our graph
        "tools": "retrieve",
        END: END,
    },
)

# Edges taken after the `action` node is called.
workflow.add_conditional_edges(
    "retrieve",
    # Assess agent decision
    grade_documents,
)
workflow.add_edge("generate_answer", END)
workflow.add_edge("rewrite_question", "generate_query_or_respond")

# Compile
graph = workflow.compile()

可视化图表

from IPython.display import Image, display

display(Image(graph.get_graph().draw_mermaid_png()))

8. 运行代理式 RAG¶

for chunk in graph.stream(
    {
        "messages": [
            {
                "role": "user",
                "content": "What does Lilian Weng say about types of reward hacking?",
            }
        ]
    }
):
    for node, update in chunk.items():
        print("Update from node", node)
        update["messages"][-1].pretty_print()
        print("\n\n")

输出

Update from node generate_query_or_respond
================================== Ai Message ==================================
Tool Calls:
  retrieve_blog_posts (call_NYu2vq4km9nNNEFqJwefWKu1)
 Call ID: call_NYu2vq4km9nNNEFqJwefWKu1
  Args:
    query: types of reward hacking



Update from node retrieve
================================= Tool Message ==================================
Name: retrieve_blog_posts

(Note: Some work defines reward tampering as a distinct category of misalignment behavior from reward hacking. But I consider reward hacking as a broader concept here.)
At a high level, reward hacking can be categorized into two types: environment or goal misspecification, and reward tampering.

Why does Reward Hacking Exist?#

Pan et al. (2022) investigated reward hacking as a function of agent capabilities, including (1) model size, (2) action space resolution, (3) observation space noise, and (4) training time. They also proposed a taxonomy of three types of misspecified proxy rewards:

Let's Define Reward Hacking#
Reward shaping in RL is challenging. Reward hacking occurs when an RL agent exploits flaws or ambiguities in the reward function to obtain high rewards without genuinely learning the intended behaviors or completing the task as designed. In recent years, several related concepts have been proposed, all referring to some form of reward hacking:



Update from node generate_answer
================================== Ai Message ==================================

Lilian Weng categorizes reward hacking into two types: environment or goal misspecification, and reward tampering. She considers reward hacking as a broad concept that includes both of these categories. Reward hacking occurs when an agent exploits flaws or ambiguities in the reward function to achieve high rewards without performing the intended behaviors.