当前位置： 首页 > news >正文

网站上网络营销,国内好用的五款开源建站系统,成都住房和城乡建设厅网站,外链工具xgAI - 谈谈RAG中的查询分析#xff08;2#xff09; 大家好#xff0c;RAG中的查询分析是比较有趣的一个点#xff0c;内容丰富#xff0c;并不是一句话能聊的清楚的。今天接着上一篇#xff0c;继续探讨RAG中的查询分析#xff0c;并在功能层面和代码层面持续改进。 功…AI - 谈谈RAG中的查询分析2 大家好RAG中的查询分析是比较有趣的一个点内容丰富并不是一句话能聊的清楚的。今天接着上一篇继续探讨RAG中的查询分析并在功能层面和代码层面持续改进。 功能层面 如果用户问了一个不着边际的问题也就是和工具无关的问题那么无须调用工具直接生成答案。否则就调用工具检索本地知识库生成答案。 代码方面 考虑到在对答聊天中对话状态是如此重要所以我们可以直接使用LangChain内置的MessagesState而不用自己定义State类。 class State(TypedDict):question: strquery: Searchcontext: List[Document]answer: str上一篇的Search工具主要用于结构化输出工具本身没有实质性内容所以本篇会将retrieve作为一个工具既可以绑定到LLM也可以通过LangGraph内置的组件 ToolNode,形成一个Graph节点在收到LLM的应答之后开始执行从本地知识库语义搜索的动作最终生成一个ToolMessage。
实例代码 备注对于本文中的代码片段主体来源于LangChain官网有兴趣的读者可以去官网查看。 import os from langchain_openai import ChatOpenAI from langchain_openai import OpenAIEmbeddings from langchain_core.vectorstores import InMemoryVectorStore import bs4 from langchain import hub from langchain_community.document_loaders import WebBaseLoader from langchain_core.documents import Document from langchain_text_splitters import RecursiveCharacterTextSplitter from langgraph.graph import START, StateGraph, MessagesState from typing_extensions import List, TypedDict from langchain_core.tools import tool from langchain_core.messages import SystemMessage from langgraph.graph import END from langgraph.prebuilt import ToolNode, tools_condition# Setup environment variables for authentication os.environ[OPENAI_API_KEY] your_openai_api_key# Initialize OpenAI embeddings using a specified model embeddings OpenAIEmbeddings(modeltext-embedding-3-large)# Create an in-memory vector store to store the embeddings vector_store InMemoryVectorStore(embeddings)# Initialize the language model from OpenAI llm ChatOpenAI(modelgpt-4o-mini)# Setup the document loader for a given web URL, specifying elements to parse loader WebBaseLoader(web_paths(https://lilianweng.github.io/posts/2023-06-23-agent/,),bs_kwargsdict(parse_onlybs4.SoupStrainer(class_(post-content, post-title, post-header))), )

Load the documents from the web page

docs loader.load()# Initialize a text splitter to chunk the document text text_splitter RecursiveCharacterTextSplitter(chunk_size1000, chunk_overlap200) all_splits text_splitter.split_documents(docs)# Index the chunks in the vector store _ vector_store.add_documents(documentsall_splits)# Define a retrieval tool to get relevant documents for a query tool(response_formatcontent_and_artifact) def retrieve(query: str):Retrieve information related to a query.retrieved_docs vector_store.similarity_search(query, k2)serialized \n\n.join((fSource: {doc.metadata}\n fContent: {doc.page_content})for doc in retrieved_docs)return serialized, retrieved_docs# Step 1: Function to generate a tool call or respond based on the state def query_or_respond(state: MessagesState):Generate tool call for retrieval or respond.llm_with_tools llm.bind_tools([retrieve]) # Bind the retrieve tool to LLMresponse llm_with_tools.invoke(state[messages]) # Invoke the LLM with current messagesreturn {messages: [response]} # Return the response messages# Step 2: Execute the retrieval tool tools ToolNode([retrieve])# Step 3: Function to generate a response using retrieved content def generate(state: MessagesState):Generate answer.# Get the most recent tool messagesrecent_tool_messages []for message in reversed(state[messages]):if message.type tool:recent_tool_messages.append(message)else:breaktool_messages recent_tool_messages[::-1] # Reverse to get the original order# Create a system message with the retrieved contextdocs_content \n\n.join(doc.content for doc in tool_messages)system_message_content (You are an assistant for question-answering tasks. Use the following pieces of retrieved context to answer the question. If you dont know the answer, say that you dont know. Use three sentences maximum and keep the answer concise.\n\nf{docs_content})# Filter human and system messages for the promptconversation_messages [messagefor message in state[messages]if message.type in (human, system)or (message.type ai and not message.tool_calls)]prompt [SystemMessage(system_message_content)] conversation_messages# Invoke the LLM with the promptresponse llm.invoke(prompt)return {messages: [response]}# Build the state graph for managing message state transitions graph_builder StateGraph(MessagesState) graph_builder.add_node(query_or_respond) # Add query_or_respond node to the graph graph_builder.add_node(tools) # Add tools node to the graph graph_builder.add_node(generate) # Add generate node to the graph# Set the entry point for the state graph graph_builder.set_entry_point(query_or_respond)

Define conditional edges based on tool invocation

graph_builder.add_conditional_edges(query_or_respond,tools_condition,{END: END, tools: tools}, ) graph_builder.add_edge(tools, generate) # Define transition from tools to generate graph_builder.add_edge(generate, END) # Define transition from generate to END# Compile the graph graph graph_builder.compile()# Interact with the compiled graph using an initial input message input_message Hello for step in graph.stream({messages: [{role: user, content: input_message}]},stream_modevalues, ):step[messages][-1].pretty_print() # Print the final message# Another interaction with the graph with a different input message input_message What is Task Decomposition? for step in graph.stream({messages: [{role: user, content: input_message}]},stream_modevalues, ):step[messages][-1].pretty_print() # Print the final message代码详解 导入必要的库 import os from langchain_openai import ChatOpenAI, OpenAIEmbeddings from langchain_core.vectorstores import InMemoryVectorStore import bs4 from langchain import hub from langchain_community.document_loaders import WebBaseLoader from langchain_core.documents import Document from langchain_text_splitters import RecursiveCharacterTextSplitter from langgraph.graph import START, StateGraph, MessagesState from typing_extensions import List, TypedDict from langchain_core.tools import tool from langchain_core.messages import SystemMessage from langgraph.graph import END from langgraph.prebuilt import ToolNode, tools_condition我们首先导入了需要的库这些库提供了处理语言和存储向量的工具。 设置环境变量 os.environ[OPENAI_API_KEY] your_openai_api_key设置一些环境变量用于API的身份验证和项目配置。 初始化嵌入模型和向量存储 embeddings OpenAIEmbeddings(modeltext-embedding-3-large) vector_store InMemoryVectorStore(embeddings)我们使用OpenAI的嵌入模型来创建文本嵌入并在内存中初始化一个向量存储用于后续的向量操作。 llm ChatOpenAI(modelgpt-4o-mini)初始化GPT-4小版本的语言模型用于后续的AI对话生成。 加载和分割文档 loader WebBaseLoader(web_paths(https://lilianweng.github.io/posts/2023-06-23-agent/,),bs_kwargsdict(parse_onlybs4.SoupStrainer(class_(post-content, post-title, post-header))), ) docs loader.load()text_splitter RecursiveCharacterTextSplitter(chunk_size1000, chunk_overlap200) all_splits text_splitter.split_documents(docs)加载指定网页的内容并对页面内容进行解析和分割。分割后的文本块将用于嵌入和向量存储。 向量存储文档 _ vector_store.add_documents(documentsall_splits)将分割后的文档片段添加到向量存储中以供后续检索操作。 定义检索工具 tool(response_formatcontent_and_artifact) def retrieve(query: str):Retrieve information related to a query.retrieved_docs vector_store.similarity_search(query, k2)serialized \n\n.join((fSource: {doc.metadata}\n fContent: {doc.page_content})for doc in retrieved_docs)return serialized, retrieved_docs定义一个检索工具函数retrieve该函数可以根据查询在向量存储中进行相似性搜索并返回检索到的文档内容。 定义步骤生成工具调用或直接回复 def query_or_respond(state: MessagesState):Generate tool call for retrieval or respond.llm_with_tools llm.bind_tools([retrieve])response llm_with_tools.invoke(state[messages])return {messages: [response]}该函数根据当前的消息状态生成调用检索工具的请求或直接生成回复。 定义步骤执行检索工具 tools ToolNode([retrieve])定义一个执行检索工具的步骤。 定义步骤生成回答 def generate(state: MessagesState):Generate answer.recent_tool_messages []for message in reversed(state[messages]):if message.type tool:recent_tool_messages.append(message)else:breaktool_messages recent_tool_messages[::-1]docs_content \n\n.join(doc.content for doc in tool_messages)system_message_content (You are an assistant for question-answering tasks. Use the following pieces of retrieved context to answer the question. If you dont know the answer, say that you dont know. Use three sentences maximum and keep the answer concise.\n\nf{docs_content})conversation_messages [messagefor message in state[messages]if message.type in (human, system)or (message.type ai and not message.tool_calls)]prompt [SystemMessage(system_message_content)] conversation_messagesresponse llm.invoke(prompt)return {messages: [response]}该函数生成最后的回答。它会首先收集最近的工具消息并结合这些消息内容生成系统消息然后与现有对话消息一起作为提示最终调用LLM生成回复。 构建状态图 graph_builder StateGraph(MessagesState) graph_builder.add_node(query_or_respond) graph_builder.add_node(tools) graph_builder.add_node(generate)graph_builder.set_entry_point(query_or_respond) graph_builder.add_conditional_edges(query_or_respond,tools_condition,{END: END, tools: tools}, ) graph_builder.add_edge(tools, generate) graph_builder.add_edge(generate, END)graph graph_builder.compile()使用状态图构建器创建一个消息状态图并添加节点和条件边确定消息的流转逻辑。 与状态图进行交互 input_message Hello for step in graph.stream({messages: [{role: user, content: input_message}]},stream_modevalues, ):step[messages][-1].pretty_print()input_message What is Task Decomposition? for step in graph.stream({messages: [{role: user, content: input_message}]},stream_modevalues, ):step[messages][-1].pretty_print()我们通过给定的输入消息与状态图进行互动流式处理消息并最终打印出生成的回复。 LLM消息抓取 以上整个过程中我们都是在调用LangChain API与LLM在进行交互至于底层发送的请求细节一无所知。在某些场景下面我们还是需要去探究一下这些具体的细节这样可以有一个全面的了解。下面我们看一下具体的发送内容以上代码涉及到三个LLM交互。 交互1 提问 {messages: [[{lc: 1,type: constructor,id: [langchain,schema,messages,HumanMessage],kwargs: {content: Hello,type: human,id: da95e909-50bb-4204-8aad-4181dcccbffb}}]] }回答 {generations: [[{text: Hello! How can I assist you today?,generation_info: {finish_reason: stop,logprobs: null},type: ChatGeneration,message: {lc: 1,type: constructor,id: [langchain,schema,messages,AIMessage],kwargs: {content: Hello! How can I assist you today?,additional_kwargs: {refusal: null},response_metadata: {token_usage: {completion_tokens: 10,prompt_tokens: 44,total_tokens: 54,completion_tokens_details: {accepted_prediction_tokens: 0,audio_tokens: 0,reasoning_tokens: 0,rejected_prediction_tokens: 0},prompt_tokens_details: {audio_tokens: 0,cached_tokens: 0}},model_name: gpt-4o-mini-2024-07-18,system_fingerprint: fp_3de1288069,finish_reason: stop,logprobs: null},type: ai,id: run-611efcc9-1fe5-47e4-83fc-f42623556d93-0,usage_metadata: {input_tokens: 44,output_tokens: 10,total_tokens: 54,input_token_details: {audio: 0,cache_read: 0},output_token_details: {audio: 0,reasoning: 0}},tool_calls: [],invalid_tool_calls: []}}}]],llm_output: {token_usage: {completion_tokens: 10,prompt_tokens: 44,total_tokens: 54,completion_tokens_details: {accepted_prediction_tokens: 0,audio_tokens: 0,reasoning_tokens: 0,rejected_prediction_tokens: 0},prompt_tokens_details: {audio_tokens: 0,cached_tokens: 0}},model_name: gpt-4o-mini-2024-07-18,system_fingerprint: fp_3de1288069},run: null,type: LLMResult }交互2 提问 {messages: [[{lc: 1,type: constructor,id: [langchain,schema,messages,HumanMessage],kwargs: {content: What is Task Decomposition?,type: human,id: 6a790b36-fafd-4ff3-b293-9bb3ac9f4157}}]] }回答 {generations: [[{text: ,generation_info: {finish_reason: tool_calls,logprobs: null},type: ChatGeneration,message: {lc: 1,type: constructor,id: [langchain,schema,messages,AIMessage],kwargs: {content: ,additional_kwargs: {tool_calls: [{id: call_RClqnmrtp2sbwIbb2jHm0VeQ,function: {arguments: {\query:\Task Decomposition},name: retrieve},type: function}],refusal: null},response_metadata: {token_usage: {completion_tokens: 15,prompt_tokens: 49,total_tokens: 64,completion_tokens_details: {accepted_prediction_tokens: 0,audio_tokens: 0,reasoning_tokens: 0,rejected_prediction_tokens: 0},prompt_tokens_details: {audio_tokens: 0,cached_tokens: 0}},model_name: gpt-4o-mini-2024-07-18,system_fingerprint: fp_0705bf87c0,finish_reason: tool_calls,logprobs: null},type: ai,id: run-056b1c5a-cd5c-40cf-940c-bbf98512615d-0,tool_calls: [{name: retrieve,args: {query: Task Decomposition},id: call_RClqnmrtp2sbwIbb2jHm0VeQ,type: tool_call}],usage_metadata: {input_tokens: 49,output_tokens: 15,total_tokens: 64,input_token_details: {audio: 0,cache_read: 0},output_token_details: {audio: 0,reasoning: 0}},invalid_tool_calls: []}}}]],llm_output: {token_usage: {completion_tokens: 15,prompt_tokens: 49,total_tokens: 64,completion_tokens_details: {accepted_prediction_tokens: 0,audio_tokens: 0,reasoning_tokens: 0,rejected_prediction_tokens: 0},prompt_tokens_details: {audio_tokens: 0,cached_tokens: 0}},model_name: gpt-4o-mini-2024-07-18,system_fingerprint: fp_0705bf87c0},run: null,type: LLMResult }交互3 提问 {messages: [[{lc: 1,type: constructor,id: [langchain,schema,messages,SystemMessage],kwargs: {content: You are an assistant for question-answering tasks. Use the following pieces of retrieved context to answer the question. If you dont know the answer, say that you dont know. Use three sentences maximum and keep the answer concise.\n\nSource: {source: https://lilianweng.github.io/posts/2023-06-23-agent/}nContent: Fig. 1. Overview of a LLM-powered autonomous agent system.\nComponent One: Planning#\nA complicated task usually involves many steps. An agent needs to know what they are and plan ahead.\nTask Decomposition#\nChain of thought (CoT; Wei et al. 2022) has become a standard prompting technique for enhancing model performance on complex tasks. The model is instructed to “think step by step” to utilize more test-time computation to decompose hard tasks into smaller and simpler steps. CoT transforms big tasks into multiple manageable tasks and shed lights into an interpretation of the model’s thinking process.\n\nSource: {source: https://lilianweng.github.io/posts/2023-06-23-agent/}nContent: Tree of Thoughts (Yao et al. 2023) extends CoT by exploring multiple reasoning possibilities at each step. It first decomposes the problem into multiple thought steps and generates multiple thoughts per step, creating a tree structure. The search process can be BFS (breadth-first search) or DFS (depth-first search) with each state evaluated by a classifier (via a prompt) or majority vote.\nTask decomposition can be done (1) by LLM with simple prompting like \Steps for XYZ.\n1.\, \What are the subgoals for achieving XYZ?\, (2) by using task-specific instructions; e.g. \Write a story outline.\ for writing a novel, or (3) with human inputs.,type: system}},{lc: 1,type: constructor,id: [langchain,schema,messages,HumanMessage],kwargs: {content: What is Task Decomposition?,type: human,id: 6a790b36-fafd-4ff3-b293-9bb3ac9f4157}}]] }回答 {generations: [[{text: Task Decomposition is the process of breaking down a complicated task into smaller, more manageable steps. It often involves techniques like Chain of Thought (CoT), where the model is prompted to think step-by-step, enhancing performance on complex tasks. This approach helps to clarify the models thinking process and makes it easier to tackle difficult problems.,generation_info: {finish_reason: stop,logprobs: null},type: ChatGeneration,message: {lc: 1,type: constructor,id: [langchain,schema,messages,AIMessage],kwargs: {content: Task Decomposition is the process of breaking down a complicated task into smaller, more manageable steps. It often involves techniques like Chain of Thought (CoT), where the model is prompted to think step-by-step, enhancing performance on complex tasks. This approach helps to clarify the models thinking process and makes it easier to tackle difficult problems.,additional_kwargs: {refusal: null},response_metadata: {token_usage: {completion_tokens: 67,prompt_tokens: 384,total_tokens: 451,completion_tokens_details: {accepted_prediction_tokens: 0,audio_tokens: 0,reasoning_tokens: 0,rejected_prediction_tokens: 0},prompt_tokens_details: {audio_tokens: 0,cached_tokens: 0}},model_name: gpt-4o-mini-2024-07-18,system_fingerprint: fp_0705bf87c0,finish_reason: stop,logprobs: null},type: ai,id: run-b3565b23-18d5-439d-a87b-f836ee281d91-0,usage_metadata: {input_tokens: 384,output_tokens: 67,total_tokens: 451,input_token_details: {audio: 0,cache_read: 0},output_token_details: {audio: 0,reasoning: 0}},tool_calls: [],invalid_tool_calls: []}}}]],llm_output: {token_usage: {completion_tokens: 67,prompt_tokens: 384,total_tokens: 451,completion_tokens_details: {accepted_prediction_tokens: 0,audio_tokens: 0,reasoning_tokens: 0,rejected_prediction_tokens: 0},prompt_tokens_details: {audio_tokens: 0,cached_tokens: 0}},model_name: gpt-4o-mini-2024-07-18,system_fingerprint: fp_0705bf87c0},run: null,type: LLMResult }总结 本文通过OpenAI语言模型和自定义检索工具构建了一个智能问答系统。首先从网络上加载和分割文档内容并将其存储到向量数据库中。然后定义一个检索工具可以根据查询请求从数据库中寻找相关文档。使用状态图管理对话流程根据不同条件系统会决定是否调用检索工具或者直接生成回复。最终通过与状态图交互实现智能应答。这样一个系统大大增强了自动化问答的能力通过结合嵌入模型和语言模型能够处理更为复杂和多样化的用户查询。