网站的跳出率很高英文杭州网站建设

网站的跳出率很高,英文杭州网站建设,揭阳企业网页制作公司,wordpress单页面模板简介 本文介绍了一种深度思考RAG流水线#xff0c;通过agent驱动的架构解决复杂查询问题。系统将查询分解为多步骤计划#xff0c;执行自适应检索策略#xff0c;结合反思机制和自我评估循环#xff0c;实现真正的多步推理。与传统线性RAG相比#xff0c;该架构能处理多源…简介本文介绍了一种深度思考RAG流水线通过agent驱动的架构解决复杂查询问题。系统将查询分解为多步骤计划执行自适应检索策略结合反思机制和自我评估循环实现真正的多步推理。与传统线性RAG相比该架构能处理多源、多跳查询并通过LangGraph构建完整认知流程显著提升答案质量。文章提供完整代码实现和评估方法是构建高级RAG系统的实用指南。很多 RAG 系统失败并不是因为 LLM 不够聪明而是因为它们的架构太简单。它们试图用线性的一次性方式处理一个本质上循环、多步骤的问题。许多复杂查询需要推理、反思以及何时行动的聪明决策这与我们面对问题时如何检索信息非常相似。这正是 RAG 流水线中引入“agent 驱动行为”的用武之地。下面看看一个典型的深度思考 RAG 流水线长什么样……Deep Thinking RAG Pipeline (Created by Fareed Khan)Plan首先agent 将复杂用户查询拆解成结构化的多步骤研究计划并决定每一步使用何种工具内部文档搜索或 web 搜索。Retrieve对每一步执行自适应的多阶段检索漏斗由一个 supervisor 动态选择最佳搜索策略vector、keyword 或 hybrid。Refine使用高精度 Cross-Encoder 对初始结果进行重排并由 distiller agent 将最佳证据压缩为简洁的上下文。Reflect每一步后agent 总结当前发现并更新研究历史逐步构建对问题的累积理解。Critique随后一个 policy agent 检查这段历史策略性决策是继续下一步、遇到死胡同时修订计划还是结束。Synthesize研究完成后最终的 agent 将来自所有来源的证据综合为单一、全面且可引用的答案。在这篇文章中我们将实现完整的“深度思考 RAG 流水线”并与基础 RAG 流水线做对比展示它是如何解决复杂的 multi-hop 查询的。所有代码与理论都在我的 GitHub 仓库GitHub - FareedKhan-dev/deep-thinking-rag: A Deep Thinking RAG Pipeline to Solve Complex Queries 目录 目录环境配置知识库来源理解多源、多跳查询构建一个会失败的浅层 RAG 流水线定义中央智能体系统的 RAG 状态战略规划与查询制定使用工具感知规划器分解问题使用查询重写智能体优化检索通过元数据感知分块提升精度创建多阶段检索漏斗使用监督器动态选择策略利用混合、关键词与语义搜索进行广泛召回使用交叉编码器重排器实现高精度通过上下文蒸馏进行综合使用网络搜索增强知识自我评估与控制流策略更新并反映累积研究历史B构建用于控制流的策略智能体定义图节点定义条件边连接深度思考 RAG 机器编译与可视化迭代工作流运行深度思考流水线分析最终高质量答案并排对比评估框架与结果分析总结整个流水线使用马尔可夫决策过程MDP学习策略环境配置在开始编写 Deep RAG 流水线前我们需要打好基础因为一个生产级 AI 系统不仅仅是最终算法还包括在搭建时做出的深思熟虑的选择。我们将要实现的每个步骤都会直接影响最终系统的有效性和可靠性。当开始开发流水线并不断试错时最好把配置定义为一个简单的字典。等流程复杂起来就可以直接回到这个字典调整配置并观察对整体性能的影响。# Central Configuration Dictionary to manage all system parametersconfig { data_dir: ./data, # Directory to store raw and cleaned data vector_store_dir: ./vector_store, # Directory to persist our vector store llm_provider: openai, # The LLM provider we are using reasoning_llm: gpt-4o, # The powerful model for planning and synthesis fast_llm: gpt-4o-mini, # A faster, cheaper model for simpler tasks like the baseline RAG embedding_model: text-embedding-3-small, # The model for creating document embeddings reranker_model: cross-encoder/ms-marco-MiniLM-L-6-v2, # The model for precision reranking max_reasoning_iterations: 7, # A safeguard to prevent the agent from getting into an infinite loop top_k_retrieval: 10, # Number of documents for initial broad recall top_n_rerank: 3, # Number of documents to keep after precision reranking}这些键大都很好理解但有三个值得强调llm_provider我们使用的 LLM 提供方这里用的是 OpenAI。之所以选择 OpenAI是因为在 LangChain 中我们可以很容易地切换模型和提供方你也可以选择适合自己的比如 Ollama。reasoning_llm在整个系统里它必须是最强的因为要承担规划与综合。fast_llm用在更简单的任务上比如 baseline RAG应更快更省。接下来导入流水线会用到的库并把 API keys 设为环境变量避免在代码中暴露。import osimport reimport jsonfrom getpass import getpassfrom pprint import pprintimport uuidfrom typing importList, Dict, TypedDict, Literal, Optionaldef_set_env(var: str): ifnot os.environ.get(var): os.environ[var] getpass(fEnter your {var}: )_set_env(OPENAI_API_KEY)_set_env(LANGSMITH_API_KEY)_set_env(TAVILY_API_KEY)os.environ[LANGSMITH_TRACING] trueos.environ[LANGSMITH_PROJECT] Advanced-Deep-Thinking-RAG我们同时启用了 LangSmith 的 tracing。在一个 agentic 系统里工作流复杂且循环tracing 并非可有可无而是很重要。它帮助你可视化内部过程更容易调试 agent 的思考路径。知识库来源一个生产级 RAG 系统需要既复杂又有挑战性的知识库才能真正体现其有效性。我们将使用 NVIDIA 的 2023 年 10-K 报告链接这是一份超过百页的文件详述公司的业务运营、财务表现和风险因素披露。Sourcing the Knowledge Base (Created by Fareed Khan)首先实现一个自定义函数直接从 SEC EDGAR 数据库下载 10-K 报告解析原始 HTML并转换成干净、结构化的文本供我们的 RAG 流水线摄取。import requestsfrom bs4 import BeautifulSoupfrom langchain.docstore.document import Documentdefdownload_and_parse_10k(url, doc_path_raw, doc_path_clean): if os.path.exists(doc_path_clean): print(fCleaned 10-K file already exists at: {doc_path_clean}) return print(fDownloading 10-K filing from {url}...) headers {User-Agent: Mozilla/5.0} response requests.get(url, headersheaders) response.raise_for_status() withopen(doc_path_raw, w, encodingutf-8) as f: f.write(response.text) print(fRaw document saved to {doc_path_raw}) soup BeautifulSoup(response.content, html.parser) text for p in soup.find_all([p, div, span]): text p.get_text(stripTrue) \n\n clean_text re.sub(r\n{3,}, \n\n, text).strip() clean_text re.sub(r\s{2,}, , clean_text).strip() withopen(doc_path_clean, w, encodingutf-8) as f: f.write(clean_text) print(fCleaned text content extracted and saved to {doc_path_clean})这段代码很直观使用 beautifulsoup4 解析 HTML 并提取文本可方便地在 HTML 结构中导航获取有效信息忽略脚本或样式等无关元素。现在执行看看效果。print(Downloading and parsing NVIDIAs 2023 10-K filing...)download_and_parse_10k(url_10k, doc_path_raw, doc_path_clean)with open(doc_path_clean, r, encodingutf-8) as f: print(\n--- Sample content from cleaned 10-K ---) print(f.read(1000) ...) plaintext #### OUTPUT ####Downloading and parsing NVIDIA 2023 10-K filing...Successfully downloaded 10-K filing from https://www.sec.gov/Archives/edgar/data/1045810/000104581023000017/nvda-20230129.htmRaw document saved to ./data/nvda_10k_2023_raw.htmlCleaned text content extracted and saved to ./data/nvda_10k_2023_clean.txt# --- Sample content from cleaned 10-K ---Item 1. Business. OVERVIEW NVIDIA is the pioneer of accelerated computing. We are a full-stack computing company with a platform strategy that brings together hardware, systems, software, algorithms, libraries, and services to create unique value for the markets we serve. Our work in accelerated computing and AI is reshaping the worlds largest industries and profoundly impacting society. Founded in 1993, we started as a PC graphics chip company, inventing the graphics processing unit, or GPU. The GPU was essential for the growth of the PC gaming market and has since been repurposed to revolutionize computer graphics, high performance computing, or HPC, and AI. The programmability of our GPUs made them ...我们调用函数把内容存到 txt 文件作为后续 RAG 的上下文。运行上述代码后会自动下载并可预览样例。理解多源、多跳查询为了测试我们实现的流水线并与基础 RAG 对比我们需要一个非常复杂的查询覆盖我们所用文档的不同方面。Our Complex Query:Based on NVIDIAs 2023 10-K filing, identify their key risks related tocompetition. Then, find recent news (post-filing, from 2024) about AMDsAI chip strategy and explain how this new strategy directly addresses orexacerbates one of NVIDIAs stated risks.为什么这个查询会难倒标准 RAG 流水线Multi-Hop 推理它不能一步完成。系统必须先识别风险再找 AMD 的新闻最后把二者综合起来。多源知识所需信息在完全不同的地方。风险在内部静态文档10-K中而 AMD 的新闻是外部的需要实时 web 访问。综合与分析不是简单列出事实而是要解释“后者如何加剧前者”需要真正的综合推理。下一节我们实现一个基础 RAG 流水线看看它是如何失败的。这份完整版的大模型 AI 学习和面试资料已经上传CSDN朋友们如果需要可以微信扫描下方CSDN官方认证二维码免费领取【保证100%免费】构建一个会失败的浅层 RAG 流水线现在环境和挑战性知识库都准备好了下一步是构建一个标准的“vanilla” RAG 流水线。这很重要……先从最简单可行的方案开始然后把复杂查询在其上运行观察到底是如何以及为何失败。我们将做如下事情Shallow RAG Pipeline (Created by Fareed Khan)加载并切分文档读取清洗后的 10-K并按固定大小切片——这是常见但语义上“天真”的方法。创建 vector store对这些切片做 embedding并用 ChromaDB 建索引支持基础语义搜索。组装 RAG Chain使用 LangChain Expression LanguageLCEL把 retriever、prompt 模板和 LLM 串起来形成线性流程。演示关键失败点用我们的多跳多源查询去执行分析其不充分的回答。先加载清洗后的文档并切分。我们用 LangChain 的RecursiveCharacterTextSplitter。from langchain_community.document_loaders import TextLoaderfrom langchain.text_splitter import RecursiveCharacterTextSplitterprint(Loading and chunking the document...)loader TextLoader(doc_path_clean, encodingutf-8)documents loader.load()text_splitter RecursiveCharacterTextSplitter(chunk_size1000, chunk_overlap150)doc_chunks text_splitter.split_documents(documents)print(fDocument loaded and split into {len(doc_chunks)} chunks.) plaintext #### OUTPUT ####Loading and chunking the document...Document loaded and split into 378 chunks.有了 378 个 chunk下一步是让它们可检索——创建向量并存入数据库。我们用 ChromaDB 作为 vector store并用 OpenAI 的text-embedding-3-small作为 embedding 模型配置中已定义。from langchain_community.vectorstores import Chromafrom langchain_openai import OpenAIEmbeddingsprint(Creating baseline vector store...)embedding_function OpenAIEmbeddings(modelconfig[embedding_model])baseline_vector_store Chroma.from_documents( documentsdoc_chunks, embeddingembedding_function)baseline_retriever baseline_vector_store.as_retriever(search_kwargs{k: 3})print(fVector store created with {baseline_vector_store._collection.count()} embeddings.) plaintext #### OUTPUT ####Creating baseline vector store...Vector store created with 378 embeddings.Chroma.from_documents会组织以上过程把向量存入可检索的索引。最后用 LCEL 把它们装配成单一可运行的 RAG chain。数据流用户问题 - retriever - prompt - LLM。from langchain_core.prompts import ChatPromptTemplatefrom langchain_openai import ChatOpenAIfrom langchain_core.runnable import RunnablePassthroughfrom langchain_core.output_parsers import StrOutputParsertemplate You are an AI financial analyst. Answer the question based only on the following context:{context}Question: {question}prompt ChatPromptTemplate.from_template(template)llm ChatOpenAI(modelconfig[fast_llm], temperature0)defformat_docs(docs): return\n\n---\n\n.join(doc.page_content for doc in docs)baseline_rag_chain ( {context: baseline_retriever | format_docs, question: RunnablePassthrough()} | prompt | llm | StrOutputParser())注意第一步是个字典。context由子链生成输入问题 -baseline_retriever-format_docs而question则原样透传RunnablePassthrough。运行看看哪里会失败。from rich.console import Consolefrom rich.markdown import Markdownconsole Console()complex_query_adv Based on NVIDIAs 2023 10-K filing, identify their key risks related to competition. Then, find recent news (post-filing, from 2024) about AMDs AI chip strategy and explain how this new strategy directly addresses or exacerbates one of NVIDIAs stated risks.print(Executing complex query on the baseline RAG chain...)baseline_result baseline_rag_chain.invoke(complex_query_adv)console.print(\n--- BASELINE RAG FAILED OUTPUT ---)console.print(Markdown(baseline_result)) plaintext #### OUTPUT ####Executing complex query on the baseline RAG chain...--- BASELINE RAG FAILED OUTPUT ---Based on the provided context, NVIDIA operates in an intensely competitive semiconductorindustry and faces competition from companies like AMD. The context mentionsthat the industry is characterized by rapid technological change. However, the provided documents do not contain any specific information about AMDs recent AI chip strategy from 2024 or how it might impact NVIDIAs stated risks.可以看到三个明显问题语境不相关retriever 抓来一些“泛泛的 NVIDIA/competition/AMD”段落却没有 2024 年 AMD 策略的具体细节。信息缺失2023 年的数据不可能覆盖 2024 年事件系统没有意识到自己“缺关键信息”。无规划与工具使用把复杂问题当成简单问答不能拆分步骤也不会用 web 搜索来补齐。系统失败不是因为 LLM 笨而是因为架构过于简单。它用线性的一次性流程试图解决一个循环的多步骤问题。理解了基础 RAG 的问题后接下来开始实现深度思考的方法论看看如何解决复杂查询。定义中央智能体系统的 RAG 状态要构建推理 agent首先需要管理它的“状态”。简单 RAG chain 的每一步都是无状态的但……智能的 agent 需要“记忆”。它需要记住最初的问题、它制定的计划、以及迄今为止收集到的证据。RAG State (Created by Fareed Khan)RAGState将作为中央记忆在我们的 LangGraph 工作流中在各节点之间传递。首先定义数据结构从最基本的构件开始研究计划中的单一步骤。我们希望定义 agent 计划的原子单元。每个Step不仅要包含一个待回答的子问题还要包含其背后的理由尤其是指定要用的工具。这迫使 agent 的规划过程明确且结构化。from langchain_core.documents import Documentfrom langchain_core.pydantic_v1 import BaseModel, Fieldclass Step(BaseModel): sub_question: str Field(descriptionA specific, answerable question for this step.) justification: str Field(descriptionA brief explanation of why this step is necessary to answer the main query.) tool: Literal[search_10k, search_web] Field(descriptionThe tool to use for this step.) keywords: List[str] Field(descriptionA list of critical keywords for searching relevant document sections.) document_section: Optional[str] Field(descriptionA likely document section title (e.g., Item 1A. Risk Factors) to search within. Only for search_10k tool.)Step类基于 PydanticBaseModel为 Planner Agent 输出提供严格契约。tool: Literal[...]强制 LLM 明确在内部知识search_10k与外部知识search_web之间做出选择。这种结构化输出比解析自然语言计划可靠得多。定义了单个Step后需要一个容器保存整个步骤序列。创建Plan类它是Step对象的列表代表 agent 端到端的研究策略。class Plan(BaseModel): steps: List[Step] Field(descriptionA detailed, multi-step plan to answer the users query.)Plan类为整个研究过程提供结构。调用 Planner Agent 时我们会要求返回符合该 schema 的 JSON 对象。这样在任何检索行动前agent 的策略都是清晰、按序的且机器可读。执行计划时agent 需要记住自己学到了什么。定义PastStep字典存储每个已完成步骤的结果构成 agent 的“研究历史”或“实验手记”。class PastStep(TypedDict): step_index: int sub_question: str retrieved_docs: List[Document] summary: str该结构对 agent 的自我批判self-critique循环至关重要。每一步后我们填充这个字典并加入 state。agent 就能通过回顾这份逐步增长的摘要列表理解自己已知/未知决定是否已具备完成任务所需的信息。最后把这些拼装为主RAGState字典。它在整个图中流动包含原始问题、完整计划、过往步骤历史以及当前正在执行步骤的中间数据。class RAGState(TypedDict): original_question: str plan: Plan past_steps: List[PastStep] current_step_index: int retrieved_docs: List[Document] reranked_docs: List[Document] synthesized_context: str final_answer: strRAGState就是 agent 的“心智”。图中的每个节点接受此字典为输入并返回更新后的版本。例如plan_node会填充plan字段retrieval_node会填充retrieved_docs以此类推。这个共享、持久的状态使复杂的迭代推理成为可能这是简单 RAG 链所缺失的。准备好 agent 的记忆蓝图后开始构建第一个认知组件Planner Agent。战略规划与查询制定Strategic Planning (Created by Fareed Khan)本节分为三步工程Tool-Aware Planner构建 LLM 驱动的 agent把用户查询分解为结构化Plan并为每步选择工具。Query Rewriter创建专门 agent把 planner 的简单子问题改写为高效检索的查询。Metadata-Aware Chunking对源文档重新处理增加 section 级 metadata这是实现高精度过滤检索的关键。使用工具感知规划器分解问题Decomposing Step (Created by Fareed Khan)不能把完整问题扔给数据库指望运气。要教会 agent 将问题拆成更小、更易处理的部分。为此我们创建专门的 Planner Agent并给出非常清晰的指令prompt告诉它该如何工作。from langchain_core.prompts import ChatPromptTemplatefrom langchain_openai import ChatOpenAIfrom rich.pretty import pprint as rprintplanner_prompt ChatPromptTemplate.from_messages([ (system, You are an expert research planner. Your task is to create a clear, multi-step plan to answer a complex user query by retrieving information from multiple sources.You have two tools available:1. search_10k: Use this to search for information within NVIDIAs 2023 10-K financial filing. This is best for historical facts, financial data, and stated company policies or risks from that specific time period.2. search_web: Use this to search the public internet for recent news, competitor information, or any topic that is not specific to NVIDIAs 2023 10-K.Decompose the users query into a series of simple, sequential sub-questions. For each step, decide which tool is more appropriate.For search_10k steps, also identify the most likely section of the 10-K (e.g., Item 1A. Risk Factors, Item 7. Managements Discussion and Analysis...).It is critical to use the exact section titles found in a 10-K filing where possible.), (human, User Query: {question})])这里给 LLM 一个新 personaexpert research planner。明确告知它有两个工具search_10k、search_web以及各自适用场景——这就是“工具感知tool-aware”的部分。我们要求它输出一个能直接映射到系统能力的计划而不是模糊表述。接下来初始化 reasoning 模型并与 prompt 串接。关键是告诉 LLM 最终输出必须符合 Pydantic 的Plan类格式确保结构化、可预测。reasoning_llm ChatOpenAI(modelconfig[reasoning_llm], temperature0)planner_agent planner_prompt | reasoning_llm.with_structured_output(Plan)print(Tool-Aware Planner Agent created successfully.)print(\n--- Testing Planner Agent ---)test_plan planner_agent.invoke({question: complex_query_adv})rprint(test_plan)我们把planner_prompt通过reasoning_llm再用.with_structured_output(Plan)让 LangChain 用函数调用能力返回完全匹配Planschema 的 JSON 对象比解析纯文本可靠得多。测试输出如下#### OUTPUT ####Tool-Aware Planner Agent created successfully.--- Testing Planner Agent ---Plan(│ steps[│ │ Step(│ │ │ sub_questionWhat are the key risks related to competition as stated in NVIDIAs 2023 10-K filing?,│ │ │ justificationThis step is necessary to extract the foundational information about competitive risks directly from the source document as requested by the user.,│ │ │ toolsearch_10k,│ │ │ keywords[competition, risk factors, semiconductor industry, competitors],│ │ │ document_sectionItem 1A. Risk Factors│ │ ),│ │ Step(│ │ │ sub_questionWhat are the recent news and developments in AMDs AI chip strategy in 2024?,│ │ │ justificationThis step requires finding up-to-date, external information that is not available in the 2023 10-K filing. A web search is necessary to get the latest details on AMDs strategy.,│ │ │ toolsearch_web,│ │ │ keywords[AMD, AI chip strategy, 2024, MI300X, Instinct accelerator],│ │ │ document_sectionNone│ │ )│ ])可以看到agent 不仅给了一个清晰的Plan还正确识别出问题包含两部分第一部分答案在 10-K 中选了search_10k且正确猜测了可能的 section。第二部分是“2024 年新闻”10-K 中不可能有正确选了search_web。这说明我们的流程在“思考层面”已有希望。使用查询重写智能体优化检索目前我们有一个包含好子问题的计划。但“有哪些风险”这样的问法并不是好的检索查询太笼统。无论是向量数据库还是 web 搜索引擎更偏好具体、关键词丰富的查询。Query Rewriting Agent (Creted by Fareed Khan)为此我们构建一个小而专的 agentQuery Rewriter。它的唯一工作是把当前步骤的子问题结合已知上下文改写为更适合检索的 query。先设计 promptfrom langchain_core.output_parsers import StrOutputParserquery_rewriter_prompt ChatPromptTemplate.from_messages([ (system, You are a search query optimization expert. Your task is to rewrite a given sub-question into a highly effective search query for a vector database or web search engine, using keywords and context from the research plan.The rewritten query should be specific, use terminology likely to be found in the target source (a financial 10-K or news articles), and be structured to retrieve the most relevant text snippets.), (human, Current sub-question: {sub_question}\n\nRelevant keywords from plan: {keywords}\n\nContext from past steps:\n{past_context})])我们让这个 agent 扮演“search query optimization expert”。它接收三类信息sub_question、keywords、past_context以此构造更强的查询。初始化 agentquery_rewriter_agent query_rewriter_prompt | reasoning_llm | StrOutputParser()print(Query Rewriter Agent created successfully.)print(\n--- Testing Query Rewriter Agent ---)test_sub_q How does AMDs 2024 AI chip strategy potentially exacerbate the competitive risks identified in NVIDIAs 10-K?test_keywords [impact, threaten, competitive pressure, market share, technological change]test_past_context Step 1 Summary: NVIDIAs 10-K lists intense competition and rapid technological change as key risks. Step 2 Summary: AMD launched its MI300X AI accelerator in 2024 to directly compete with NVIDIAs H100.rewritten_q query_rewriter_agent.invoke({ sub_question: test_sub_q, keywords: test_keywords, past_context: test_past_context})print(fOriginal sub-question: {test_sub_q})print(fRewritten Search Query: {rewritten_q})结果如下#### OUTPUT ####Query Rewriter Agent created successfully.--- Testing Query Rewriter Agent ---Original sub-question: How does AMD 2024 AI chip strategy potentially exacerbate the competitive risks identified in NVIDIA 10-K?Rewritten Search Query: analysis of how AMD 2024 AI chip strategy, including products like the MI300X, exacerbates NVIDIAs stated competitive risks such as rapid technological change and market share erosion in the data center and AI semiconductor industry原问题面向分析师改写后的查询面向搜索引擎包含更具体术语如“MI300X”、“market share erosion”、“data center”等这些都从关键词和过往上下文中归纳出来。这样的查询更可能检回准确文档提升系统准确与效率。通过元数据感知分块提升精度Planner Agent 让我们有了“额外线索”它不仅说“找风险”还提示“看 Item 1A. Risk Factors 这一节”。但当前 retriever 用不上这个提示。vector store 只是 378 个 chunk 的“扁平列表”不知道什么是“section”。Meta aware chunking (Created by Fareed Khan)我们需要重建 chunks。这次每个 chunk 都要加上“它来自 10-K 的哪一节”的 metadata 标签。这样 agent 就能执行更精确的“过滤检索”。首先需要在原始文本中程序化定位每个 section 的起始。观察文档格式每个主 section 以 “ITEM”编号开头如“ITEM 1A”、“ITEM 7”非常适合用正则。section_pattern r(ITEM\\s\\d[A-Z]?\\.\\s*.*?)(?\\nITEM\\s\\d[A-Z]?\\.|$)这条 pattern 用于检测 section 标题既要足够灵活以适配多种格式又要足够具体避免误抓。应用该 pattern把文档切分为两个列表section 标题列表、对应内容列表。raw_text documents[0].page_contentsection_titles re.findall(section_pattern, raw_text, re.IGNORECASE | re.DOTALL)section_titles [title.strip().replace(\\n, ) for title in section_titles]sections_content re.split(section_pattern, raw_text, flagsre.IGNORECASE | re.DOTALL)sections_content [content.strip() for content in sections_content if content.strip() and not content.strip().lower().startswith(item )]print(fIdentified {len(section_titles)} document sections.)assert len(section_titles) len(sections_content), Mismatch between titles and content sections这是一种高效解析半结构化文档的方法。用一次findall获得所有 section 标题再用一次split按标题切分全文。assert是健全性检查确保解析正确。接着将标题与内容逐一对应生成最终带 metadata 的 chunks。import uuiddoc_chunks_with_metadata []for i, content inenumerate(sections_content): section_title section_titles[i] section_chunks text_splitter.split_text(content) for chunk in section_chunks: chunk_id str(uuid.uuid4()) doc_chunks_with_metadata.append( Document( page_contentchunk, metadata{ section: section_title, source_doc: doc_path_clean, id: chunk_id } ) )print(fCreated {len(doc_chunks_with_metadata)} chunks with section metadata.)print(\n--- Sample Chunk with Metadata ---)sample_chunk next(c for c in doc_chunks_with_metadata ifRisk Factorsin c.metadata.get(section, ))print(sample_chunk)核心在于为每个 chunk 附加metadata将section_title作为标签。输出如下#### OUTPUT ####Processing document and adding metadata...Identified 22 document sections.Created 381 chunks with section metadata.--- Sample Chunk with Metadata ---Document(│ page_contentOur industry is intensely competitive. We operate in the semiconductor\\nindustry, which is intensely competitive and characterized by rapid\\ntechnological change and evolving industry standards. ...│ metadata{│ │ section: Item 1A. Risk Factors.,│ │ source_doc: ./data/nvda_10k_2023_clean.txt,│ │ id: ...│ })看到metadata中的section: Item 1A. Risk Factors.了吗现在当 agent 需要找“风险”时可以对 retriever 说“只在 section‘Item 1A. Risk Factors’ 的 chunks 中检索”。一个简单的改造就让检索从“钝器”变成“手术刀”。这是构建生产级 RAG 的关键原则。创建多阶段检索漏斗到目前为止我们已经做了智能规划并为文档添加了 metadata。现在构建系统的核心复杂的检索流水线。简单的一次性语义搜索已经不够。生产级 agent 需要自适应、分阶段的检索流程。Multi Stage Funnel (Created by Fareed Khan)Retrieval Supervisor构建 supervisor agent 作为动态路由器分析每个子问题并选择最佳检索策略vector、keyword 或 hybrid。第一阶段广覆盖 Recall实现 supervisor 可选的不同检索策略尽可能广泛地捕获潜在相关文档。第二阶段高精度 Precision使用 Cross-Encoder 模型对初始结果进行重排去噪并将最相关文档置顶。第三阶段综合 Synthesis创建 Distiller Agent 将 top 文档压缩为单一、简洁的上下文。使用监督器动态选择策略并非所有查询都相同。比如“Compute Networking 分部 2023 财年的 revenue 增长是多少”包含非常具体的术语keyword 搜索更合适而“公司对市场竞争的整体态度如何”则是概念性问题semantic 搜索更优。Supervisor Agent (Created by Fareed Khan)我们不硬编码策略而是构建一个小而智能的 agent——Retrieval Supervisor。它的职责就是分析查询决定用哪种检索方式。先定义其输出的结构class RetrievalDecision(BaseModel): strategy: Literal[vector_search, keyword_search, hybrid_search] justification: str然后是 promptretrieval_supervisor_prompt ChatPromptTemplate.from_messages([ (system, You are a retrieval strategy expert. Based on the users query, you must decide the best retrieval strategy.You have three options:1. vector_search: Best for conceptual, semantic, or similarity-based queries.2. keyword_search: Best for queries with specific, exact terms, names, or codes (e.g., Item 1A, Hopper architecture).3. hybrid_search: A good default that combines both, but may be less precise than a targeted strategy.), (human, User Query: {sub_question})])装配该 agent 并测试retrieval_supervisor_agent retrieval_supervisor_prompt | reasoning_llm.with_structured_output(RetrievalDecision)print(Retrieval Supervisor Agent created.)print(\n--- Testing Retrieval Supervisor Agent ---)query1 revenue growth for the Compute Networking segment in fiscal year 2023decision1 retrieval_supervisor_agent.invoke({sub_question: query1})print(fQuery: {query1})print(fDecision: {decision1.strategy}, Justification: {decision1.justification})query2 general sentiment about market competition and technological innovationdecision2 retrieval_supervisor_agent.invoke({sub_question: query2})print(f\nQuery: {query2})print(fDecision: {decision2.strategy}, Justification: {decision2.justification}) plaintext #### OUTPUT ####Retrieval Supervisor Agent created.# --- Testing Retrieval Supervisor Agent ---Query: revenue growth for the Compute Networking segment in fiscal year 2023Decision: keyword_search, Justification: The query contains specific keywords like revenue growth, Compute Networking, and fiscal year 2023 which are ideal for a keyword-based search to find exact financial figures.Query: general sentiment about market competition and technological innovationDecision: vector_search, Justification: This query is conceptual and seeks to understand sentiment and broader themes. Vector search is better suited to capture the semantic meaning of market competition and technological innovation rather than relying on exact keywords.它能正确地为具体术语选keyword_search为概念性问题选vector_search。动态决策比一刀切强得多。利用混合、关键词与语义搜索进行广泛召回有了 supervisor 选择策略我们需要实现这些策略。第一阶段的目标是 Recall广覆盖尽可能捕获所有潜在相关文档即使带入一些噪声也没关系。Broad Recall (Created by Fareed Khan)我们实现三种搜索函数Vector Search标准语义搜索升级为支持 metadata filter。Keyword SearchBM25传统且强大的算法擅长匹配具体术语。Hybrid Search结合二者用 RRFReciprocal Rank Fusion融合。先用带 metadata 的 chunks 创建一个高级 vector store。import numpy as npfrom rank_bm25 import BM25Okapiprint(Creating advanced vector store with metadata...)advanced_vector_store Chroma.from_documents( documentsdoc_chunks_with_metadata, embeddingembedding_function)print(fAdvanced vector store created with {advanced_vector_store._collection.count()} embeddings.)接着构建 BM25 的索引print(\nBuilding BM25 index for keyword search...)tokenized_corpus [doc.page_content.split( ) for doc in doc_chunks_with_metadata]doc_ids [doc.metadata[id] for doc in doc_chunks_with_metadata]doc_map {doc.metadata[id]: doc for doc in doc_chunks_with_metadata}bm25 BM25Okapi(tokenized_corpus)定义三个检索函数def vector_search_only(query: str, section_filter: str None, k: int 10): filter_dict {section: section_filter} if section_filter andUnknownnotin section_filter elseNone return advanced_vector_store.similarity_search(query, kk, filterfilter_dict)defbm25_search_only(query: str, k: int 10): tokenized_query query.split( ) bm25_scores bm25.get_scores(tokenized_query) top_k_indices np.argsort(bm25_scores)[::-1][:k] return [doc_map[doc_ids[i]] for i in top_k_indices]defhybrid_search(query: str, section_filter: str None, k: int 10): bm25_docs bm25_search_only(query, kk) semantic_docs vector_search_only(query, section_filtersection_filter, kk) all_docs {doc.metadata[id]: doc for doc in bm25_docs semantic_docs}.values() ranked_lists [[doc.metadata[id] for doc in bm25_docs], [doc.metadata[id] for doc in semantic_docs]] rrf_scores {} for doc_list in ranked_lists: for i, doc_id inenumerate(doc_list): if doc_id notin rrf_scores: rrf_scores[doc_id] 0 rrf_scores[doc_id] 1 / (i 61) sorted_doc_ids sorted(rrf_scores.keys(), keylambda x: rrf_scores[x], reverseTrue) final_docs [doc_map[doc_id] for doc_id in sorted_doc_ids[:k]] return final_docsprint(\nAll retrieval strategy functions ready.)快速测试 keyword 搜索是否能精确命中目标 sectionprint(\n--- Testing Keyword Search ---)test_query Item 1A. Risk Factorstest_results bm25_search_only(test_query)print(fQuery: {test_query})print(fFound {len(test_results)} documents. Top result section: {test_results[0].metadata[section]}) plaintext #### OUTPUT ####Creating advanced vector store with metadata...Advanced vector store created with 381 embeddings.Building BM25 index for keyword search...All retrieval strategy functions ready.# --- Testing Keyword Search ---Query: Item 1A. Risk FactorsFound 10 documents. Top result section: Item 1A. Risk Factors.如预期BM25 能精确、迅速地检回 “Item 1A. Risk Factors” 相关文档。当查询包含具体关键词如 section 标题时supervisor 就可以选择这一精准工具。接下来进入高精度阶段进行重排。这份完整版的大模型 AI 学习和面试资料已经上传CSDN朋友们如果需要可以微信扫描下方CSDN官方认证二维码免费领取【保证100%免费】使用交叉编码器重排器实现高精度第一阶段的 Recall 能拿到 10 个“潜在相关”的文档。但“潜在相关”还不够直接把这 10 个 chunk 全喂给主推理 LLM 会既低效又有风险——成本高还可能被噪声干扰。High Precision (Created by Fareed Khan)我们需要 Precision 阶段用 Reranker 来从这 10 个候选中挑出最相关的少数。区别在于模型工作方式初始检索用的是 bi-encoderembedding 模型分别对 query 与文档编码速度快、适合海量搜索。Cross-Encoder 则将“query 单个文档”作为一对一起输入做更深入的比较。它更慢但更准。我们要写一个函数把 10 个文档用 Cross-Encoder 打分重排只保留config里的 top 3。初始化 Cross-Encoder 模型from sentence_transformers import CrossEncoderprint(Initializing CrossEncoder reranker...)reranker CrossEncoder(config[reranker_model])定义重排函数def rerank_documents_function(query: str, documents: List[Document]) - List[Document]: if not documents: return [] pairs [(query, doc.page_content) for doc in documents] scores reranker.predict(pairs) doc_scores list(zip(documents, scores)) doc_scores.sort(keylambda x: x[1], reverseTrue) reranked_docs [doc for doc, score in doc_scores[:config[top_n_rerank]]] return reranked_docs该函数用 cross-encoder 对每个query, doc对进行打分排序后截取前 3输出短小而高相关的文档列表作为后续 agent 的完美上下文。这样的漏斗式处理先高召回再高精度是生产级 RAG 的关键。通过上下文蒸馏进行综合现在我们有 10 - 3 的高相关文档但仍然是三个独立块。为进一步精炼再加入最后一道“压缩”——Contextual Distillation将前 3 个文档“蒸馏”为一个简洁、干净的段落去除冗余构建一段信息密度极高的上下文。Synthesization (Created by Fareed Khan)这一步是针对文本处理不负责回答问题。我们创建 Distiller Agentdistiller_prompt ChatPromptTemplate.from_messages([ (system, You are a helpful assistant. Your task is to synthesize the following retrieved document snippets into a single, concise paragraph.The goal is to provide a clear and coherent context that directly answers the question: {question}.Focus on removing redundant information and organizing the content logically. Answer only with the synthesized context.), (human, Retrieved Documents:\n{context})])distiller_agent distiller_prompt | reasoning_llm | StrOutputParser()print(Contextual Distiller Agent created.)在主循环中每一步的流程将是Supervisor选择检索策略vector/keyword/hybrid。Recall执行选择的策略取 top 10 文档。Precision用rerank_documents_function取 top 3。Distillation用distiller_agent压缩为单段精华。这样我们的证据质量达到最佳。下一步给 agent “看向外部世界”的能力web 检索。使用网络搜索增强知识目前的检索漏斗很强但有一个致命盲点它只能看到 2023 年 10-K 中的内容。而我们的挑战需要“2024 年 AMD 的 AI 芯片策略”的最新新闻——这些在静态知识库中根本不存在。真正的“深度思考” agent必须意识到自身知识的边界并能到别处找答案。我们需要给它一扇“窗”。Augemtation using Web (Created by Fareed Khan)这一步我们为系统增加一个新工具Web Search使其从“文档特定问答机器人”变成真正的多源研究助手。我们使用 Tavily Search API——专为 LLM 构建的搜索引擎返回干净、无广告、相关的结果非常适合 RAG同时与 LangChain 集成顺畅。初始化 Tavily 搜索工具from langchain_community.tools.tavily_search import TavilySearchResultsweb_search_tool TavilySearchResults(k3)原始 API 响应需要包装为标准的Document列表以便与我们的 reranker、distiller 无缝衔接。写一个小包装函数def web_search_function(query: str) - List[Document]: results web_search_tool.invoke({query: query}) return [ Document( page_contentres[content], metadata{source: res[url]} ) for res in results ]测试print(\n--- Testing Web Search Tool ---)test_query_web AMD AI chip strategy 2024test_results_web web_search_function(test_query_web)print(fFound {len(test_results_web)} results for query: {test_query_web})if test_results_web: print(fTop result snippet: {test_results_web[0].page_content[:250]}...) plaintext #### OUTPUT ####Web search tool (Tavily) initialized.--- Testing Web Search Tool ---Found 3 results for query: AMD AI chip strategy 2024Top result snippet: AMD has intensified its battle with Nvidia in the AI chip market with the release of the Instinct MI300X accelerator, a powerful GPU designed to challenge Nvidias H100 in training and inference for large language models. Major cloud providers like Microsoft Azure and Oracle Cloud are adopting the MI300X, indicating strong market interest...结果如愿找到了 3 篇相关网页。摘要提到了 AMD “Instinct MI300X” 与 NVIDIA “H100”的对抗——正是解决第二部分问题所需的证据。现在 agent 拥有通往外部世界的窗口planner 可以智能地决定何时使用它。下一步是让 agent 能够“反思”并决定何时结束研究。自我评估与控制流策略到目前为止agent 能制定计划、选择工具并执行复杂的检索漏斗。但还缺少一个关键能力对自身进展进行“思考”。盲目照搬计划逐步执行的 agent 并非真正智能需要一个自我批判机制。Self Critique and Policy Making (Created by Fareed Khan)每次研究步骤后agent 都会停下来反思比较新信息与既有知识然后做出策略性决策研究是否已完成还是要继续这个自我批判循环让系统从脚本化工作流跃升为自治 agent。它能判断自己是否已经收集到足够的证据来有信心地回答用户问题。我们将实现两个专门 agentReflection Agent读取当前步骤的精炼上下文写一条简洁的一句话摘要加入“研究历史”。Policy Agent作为总指挥在反思之后审视整个历史与最初计划做出关键决策CONTINUE_PLAN或FINISH。更新并反映累积研究历史每完成一步例如检索并蒸馏出 NVIDIA 的风险不要直接进入下一步。需要把新知识整合到 agent 的记忆中。Reflective Cumulative (Created by Fareed Khan)构建 Reflection Agent任务是读入当前步骤的精炼上下文写一条事实性的一句话摘要并把它添加到RAGState的past_steps中。reflection_prompt ChatPromptTemplate.from_messages([ (system, You are a research assistant. Based on the retrieved context for the current sub-question, write a concise, one-sentence summary of the key findings.This summary will be added to our research history. Be factual and to the point.), (human, Current sub-question: {sub_question}\n\nDistilled context:\n{context})])reflection_agent reflection_prompt | reasoning_llm | StrOutputParser()print(Reflection Agent created.)它是认知循环的重要组成通过这些简洁摘要构建干净易读的“研究历史”为下一个、也是最重要的 agent——策略决策者提供输入。B构建用于控制流的策略智能体这是 agent 自主性的“大脑”。在reflection_agent更新历史后Policy Agent 上场作为总调度查看原始问题、初始计划、已完成步骤摘要的全量历史做出高阶策略决策。Policy Agent (Created by Fareed Khan)定义决策输出结构class Decision(BaseModel): next_action: Literal[CONTINUE_PLAN, FINISH] justification: str设计 promptpolicy_prompt ChatPromptTemplate.from_messages([ (system, You are a master strategist. Your role is to analyze the research progress and decide the next action.You have the original question, the initial plan, and a log of completed steps with their summaries.- If the collected information in the Research History is sufficient to comprehensively answer the Original Question, decide to FINISH.- Otherwise, if the plan is not yet complete, decide to CONTINUE_PLAN.), (human, Original Question: {question}\n\nInitial Plan:\n{plan}\n\nResearch History (Completed Steps):\n{history})])policy_agent policy_prompt | reasoning_llm.with_structured_output(Decision)print(Policy Agent created.)测试两个状态print(\n--- Testing Policy Agent (Incomplete State) ---)plan_str json.dumps([s.dict() for s in test_plan.steps])incomplete_history Step 1 Summary: NVIDIAs 10-K states that the semiconductor industry is intensely competitive and subject to rapid technological change.decision1 policy_agent.invoke({question: complex_query_adv, plan: plan_str, history: incomplete_history})print(fDecision: {decision1.next_action}, Justification: {decision1.justification})print(\n--- Testing Policy Agent (Complete State) ---)complete_history incomplete_history \nStep 2 Summary: In 2024, AMD launched its MI300X accelerator to directly compete with NVIDIA in the AI chip market, gaining adoption from major cloud providers.decision2 policy_agent.invoke({question: complex_query_adv, plan: plan_str, history: complete_history})print(fDecision: {decision2.next_action}, Justification: {decision2.justification}) plaintext #### OUTPUT ####Policy Agent created.--- Testing Policy Agent (Incomplete State) ---Decision: CONTINUE_PLAN, Justification: The research has only identified NVIDIAs competitive risks from the 10-K. It has not yet gathered the required external information about AMDs 2024 strategy, which is the next step in the plan.--- Testing Policy Agent (Complete State) ---Decision: FINISH, Justification: The research history now contains comprehensive summaries of both NVIDIAs stated competitive risks and AMDs recent AI chip strategy. All necessary information has been gathered to perform the final synthesis and answer the users question.未完成状态时正确选择CONTINUE_PLAN完成状态时正确选择FINISH。有了policy_agent我们具备自主系统的头脑。接下来用 LangGraph 把所有组件串起来。这份完整版的大模型 AI 学习和面试资料已经上传CSDN朋友们如果需要可以微信扫描下方CSDN官方认证二维码免费领取【保证100%免费】定义图节点我们已经设计好这些专门的 agent。现在要把它们变成工作流的“积木”。LangGraph 中的“节点node”就是干这事的每个节点是一个 Python 函数完成一项具体工作接收RAGState更新并返回字典。Graph Nodes (Created by Fareed Khan)先写一个工具函数把研究历史past_steps格式化成易读字符串方便传给 promptdef get_past_context_str(past_steps: List[PastStep]) - str: return \\n\\n.join([fStep {s[step_index]}: {s[sub_question]}\\nSummary: {s[summary]} for s in past_steps])第一个节点plan_node调用planner_agent填充plan字段。def plan_node(state: RAGState) - Dict: console.print(--- : Generating Plan ---) plan planner_agent.invoke({question: state[original_question]}) rprint(plan) return {plan: plan, current_step_index: 0, past_steps: []}接着是两个检索节点内部文档与 web。def retrieval_node(state: RAGState) - Dict: current_step_index state[current_step_index] current_step state[plan].steps[current_step_index] console.print(f--- : Retrieving from 10-K (Step {current_step_index 1}: {current_step.sub_question}) ---) past_context get_past_context_str(state[past_steps]) rewritten_query query_rewriter_agent.invoke({ sub_question: current_step.sub_question, keywords: current_step.keywords, past_context: past_context }) console.print(f Rewritten Query: {rewritten_query}) retrieval_decision retrieval_supervisor_agent.invoke({sub_question: rewritten_query}) console.print(f Supervisor Decision: Use {retrieval_decision.strategy}. Justification: {retrieval_decision.justification}) if retrieval_decision.strategy vector_search: retrieved_docs vector_search_only(rewritten_query, section_filtercurrent_step.document_section, kconfig[top_k_retrieval]) elif retrieval_decision.strategy keyword_search: retrieved_docs bm25_search_only(rewritten_query, kconfig[top_k_retrieval]) else: retrieved_docs hybrid_search(rewritten_query, section_filtercurrent_step.document_section, kconfig[top_k_retrieval]) return {retrieved_docs: retrieved_docs} plaintext def web_search_node(state: RAGState) - Dict: current_step_index state[current_step_index] current_step state[plan].steps[current_step_index] console.print(f--- : Searching Web (Step {current_step_index 1}: {current_step.sub_question}) ---) past_context get_past_context_str(state[past_steps]) rewritten_query query_rewriter_agent.invoke({ sub_question: current_step.sub_question, keywords: current_step.keywords, past_context: past_context }) console.print(f Rewritten Query: {rewritten_query}) retrieved_docs web_search_function(rewritten_query) return {retrieved_docs: retrieved_docs}然后是 Precision 与 Distillation 节点def rerank_node(state: RAGState) - Dict: console.print(--- : Reranking Documents ---) current_step_index state[current_step_index] current_step state[plan].steps[current_step_index] reranked_docs rerank_documents_function(current_step.sub_question, state[retrieved_docs]) console.print(f Reranked to top {len(reranked_docs)} documents.) return {reranked_docs: reranked_docs} plaintext def compression_node(state: RAGState) - Dict: console.print(--- ✂️: Distilling Context ---) current_step_index state[current_step_index] current_step state[plan].steps[current_step_index] context format_docs(state[reranked_docs]) synthesized_context distiller_agent.invoke({question: current_step.sub_question, context: context}) console.print(f Distilled Context Snippet: {synthesized_context[:200]}...) return {synthesized_context: synthesized_context}反思并更新历史def reflection_node(state: RAGState) - Dict: console.print(--- : Reflecting on Findings ---) current_step_index state[current_step_index] current_step state[plan].steps[current_step_index] summary reflection_agent.invoke({sub_question: current_step.sub_question, context: state[synthesized_context]}) console.print(f Summary: {summary}) new_past_step { step_index: current_step_index 1, sub_question: current_step.sub_question, retrieved_docs: state[reranked_docs], summary: summary } return {past_steps: state[past_steps] [new_past_step], current_step_index: current_step_index 1}最终答案生成def final_answer_node(state: RAGState) - Dict: console.print(--- ✅: Generating Final Answer with Citations ---) final_context for i, step inenumerate(state[past_steps]): final_context f\\n--- Findings from Research Step {i1} ---\\n for doc in step[retrieved_docs]: source doc.metadata.get(section) or doc.metadata.get(source) final_context fSource: {source}\\nContent: {doc.page_content}\\n\\n final_answer_prompt ChatPromptTemplate.from_messages([ (system, You are an expert financial analyst. Synthesize the research findings from internal documents and web searches into a comprehensive, multi-paragraph answer for the users original question.Your answer must be grounded in the provided context. At the end of any sentence that relies on specific information, you MUST add a citation. For 10-K documents, use [Source: ]. For web results, use [Source: ].), (human, Original Question: {question}\n\nResearch History and Context:\n{context}) ]) final_answer_agent final_answer_prompt | reasoning_llm | StrOutputParser() final_answer final_answer_agent.invoke({question: state[original_question], context: final_context}) return {final_answer: final_answer}节点齐备后接下来定义“边”edges确定它们之间的连接关系与控制流。定义条件边我们需要两类关键的条件边工具路由器route_by_tool在plan之后查看当前步骤应使用的工具路由到retrieve_10k或retrieve_web。主控制循环should_continue_node每次反思后调用policy_agent决定是继续下一步还是结束并生成答案。工具路由器def route_by_tool(state: RAGState) - str: current_step_index state[current_step_index] current_step state[plan].steps[current_step_index] return current_step.tool主控制循环def should_continue_node(state: RAGState) - str: console.print(--- : Evaluating Policy ---) current_step_index state[current_step_index] if current_step_index len(state[plan].steps): console.print( - Plan complete. Finishing.) returnfinish if current_step_index config[max_reasoning_iterations]: console.print( - Max iterations reached. Finishing.) returnfinish if state.get(reranked_docs) isnotNoneandnot state[reranked_docs]: console.print( - Retrieval failed for the last step. Continuing with next step in plan.) returncontinue history get_past_context_str(state[past_steps]) plan_str json.dumps([s.dict() for s in state[plan].steps]) decision policy_agent.invoke({question: state[original_question], plan: plan_str, history: history}) console.print(f - Decision: {decision.next_action} | Justification: {decision.justification}) if decision.next_action FINISH: returnfinish else: returncontinue有了节点专家与条件边对话规则我们就可以构建完整的StateGraph。连接深度思考 RAG 机器现在用 LangGraph 的StateGraph来定义完整的认知架构也就是 agent 的思维流程蓝图。from langgraph.graph import StateGraph, ENDgraph StateGraph(RAGState)添加节点graph.add_node(plan, plan_node)graph.add_node(retrieve_10k, retrieval_node)graph.add_node(retrieve_web, web_search_node)graph.add_node(rerank, rerank_node)graph.add_node(compress, compression_node)graph.add_node(reflect, reflection_node)graph.add_node(generate_final_answer, final_answer_node)连接边与条件边graph.set_entry_point(plan)graph.add_conditional_edges( plan, route_by_tool, { search_10k: retrieve_10k, search_web: retrieve_web, },)graph.add_edge(retrieve_10k, rerank)graph.add_edge(retrieve_web, rerank)graph.add_edge(rerank, compress)graph.add_edge(compress, reflect)graph.add_conditional_edges( reflect, should_continue_node, { continue: plan, finish: generate_final_answer, },)graph.add_edge(generate_final_answer, END)print(StateGraph constructed successfully.)流程回顾从plan开始route_by_tool决定走retrieve_10k还是retrieve_web然后始终按rerank - compress - reflectreflect后通过should_continue_node决定若CONTINUE_PLAN回到plan路由下一步若FINISH进入generate_final_answer生成最终答案后结束。至此我们完成了深度思考 Agent 的复杂、循环架构。下一步是编译与可视化。编译与可视化迭代工作流编译.compile()会把抽象的节点与边定义转化为可执行应用。我们还可以用内置工具生成图示有助于理解与调试。deep_thinking_rag_graph graph.compile()print(Graph compiled successfully.)try: from IPython.display import Image, display png_image deep_thinking_rag_graph.get_graph().draw_png() display(Image(png_image))except Exception as e: print(fGraph visualization failed: {e}. Please ensure pygraphviz is installed.)  Deep Thinking Simpler Pipeline Flow (Created by Fareed Khan) 你会看到 * route_by_tool 选择内部或外部检索的分支 * 每个研究步骤的线性处理rerank - compress - reflect * 关键的反馈循环should_continue 把流程送回 plan开始下一轮 * 研究完成后进入 generate_final_answer 的“出口”。 这就是一个“会思考”的系统。接下来实际运行。 运行深度思考流水线 --------- 我们要用相同的多跳多源查询来测试这个系统看看它能否成功。 这里我们调用 .stream() 观察每个节点更新后的 state实时追踪 agent 的“思考过程”。 plaintext final_state Nonegraph_input {original_question: complex_query_adv}print(--- Invoking Deep Thinking RAG Graph ---)for chunk in deep_thinking_rag_graph.stream(graph_input, stream_modevalues): final_state chunkprint(\n--- Graph Stream Finished ---) plaintext #### OUTPUT ####--- Invoking Deep Thinking RAG Graph ------ : Generating Plan ---plan: steps: - sub_question: What are the key risks related to competition as stated in NVIDIAs 2023 10-K filing? tool: search_10k ... - sub_question: What are the recent news and developments in AMDs AI chip strategy in 2024? tool: search_web ...--- : Retrieving from 10-K (Step 1: ...) --- Rewritten Query: key competitive risks for NVIDIA in the semiconductor industry... Supervisor Decision: Use hybrid_search. ...--- : Reranking Documents --- Reranked to top 3 documents.--- ✂️: Distilling Context --- Distilled Context Snippet: NVIDIA operates in the intensely competitive semiconductor industry...--- : Reflecting on Findings --- Summary: According to its 2023 10-K, NVIDIA operates in an intensely competitive semiconductor industry...--- : Evaluating Policy --- - Decision: CONTINUE_PLAN | Justification: The first step...has been completed. The next step...is still pending...--- : Searching Web (Step 2: ...) --- Rewritten Query: AMD AI chip strategy news and developments 2024...--- : Reranking Documents --- Reranked to top 3 documents.--- ✂️: Distilling Context --- Distilled Context Snippet: AMD has ramped up its challenge to Nvidia in the AI accelerator market with its Instinct MI300 series...--- : Reflecting on Findings --- Summary: In 2024, AMD is aggressively competing with NVIDIA in the AI chip market through its Instinct MI300X accelerator...--- : Evaluating Policy --- - Decision: FINISH | Justification: The research history now contains comprehensive summaries of both NVIDIAs stated risks and AMDs recent strategy...--- ✅: Generating Final Answer with Citations ------ Graph Stream Finished ---可以看到系统完整执行了我们设计的流程规划 - 步骤 1 - 自我评估继续 - 步骤 2 - 自我评估结束 - 最终综合。分析最终高质量答案打印最终答案console.print(--- DEEP THINKING RAG FINAL ANSWER ---)console.print(Markdown(final_state[final_answer])) plaintext #### OUTPUT ####--- DEEP THINKING RAG FINAL ANSWER ---Based on an analysis of NVIDIAs 2023 10-K filing and recent news from 2024 regarding AMDs AI chip strategy, the following synthesis can be made:**NVIDIAs Stated Competitive Risks:**In its 2023 10-K filing, NVIDIA identifies its operating environment as the intensely competitive semiconductor industry, which is characterized by rapid technological change. A primary risk is that competitors, including AMD, could introduce new products with better performance or lower costs that gain significant market acceptance, which could materially and adversely affect its business [Source: Item 1A. Risk Factors.].**AMDs 2024 AI Chip Strategy:**In 2024, AMD has moved aggressively to challenge NVIDIAs dominance in the AI hardware market with its Instinct MI300 series of accelerators, particularly the MI300X. This product is designed to compete directly with NVIDIAs H100 GPU. AMDs strategy has gained significant traction, with major cloud providers such as Microsoft Azure and Oracle announcing plans to use the new chips [Source: https://www.reuters.com/technology/amd-forecasts-35-billion-ai-chip-revenue-2024-2024-01-30/].**Synthesis and Impact:**AMDs 2024 AI chip strategy directly exacerbates the competitive risks outlined in NVIDIAs 10-K. The successful launch and adoption of the MI300X is a materialization of the specific risk that a competitor could introduce a product with comparable performance. The adoption of AMDs chips by major cloud providers signifies a direct challenge to NVIDIAs market share in the lucrative data center segment, validating NVIDIAs stated concerns about rapid technological change [Source: Item 1A. Risk Factors. and https://www.cnbc.com/2023/12/06/amd-launches-new-mi300x-ai-chip-to-compete-with-nvidias-h100.html].这是一次“完全成功”的综合性回答正确总结了 10-K 的风险正确总结了 2024 年 AMD 动向关键在“综合与影响”部分完成了多跳推理解释“后者如何加剧前者”并提供了来源溯源内部 section 与外部 URL。并排对比让我们把两种结果放在一起对比。这个对比清晰地说明采用循环、工具感知、自我批判的 agent 架构在复杂真实查询上实现了显著且可量化的性能提升。评估框架与结果分析虽然我们在一个难题上取得了成功但在生产环境中需要客观、量化、自动化的验证。Evaluation Framework (Created by Fareed Khan)我们使用 RAGAsRAG Assessment库聚焦四个关键指标Context Precision Recall衡量检索质量。Precision 问检回的文档有多少真相关Recall 问所有相关文档中我们找到了多少Answer Faithfulness答案是否扎根于提供的上下文是防止 LLM 幻觉的主要检查。Answer Correctness最终质量度量与人工撰写的“ground truth”答案对比评估事实准确性与完整性。准备评估数据集包含问题、两套系统的答案与上下文、以及 ground truth并评测from datasets import Datasetfrom ragas import evaluatefrom ragas.metrics import ( context_precision, context_recall, faithfulness, answer_correctness,)import pandas as pdprint(Preparing evaluation dataset...)ground_truth_answer_adv NVIDIAs 2023 10-K lists intense competition and rapid technological change as key risks. This risk is exacerbated by AMDs 2024 strategy, specifically the launch of the MI300X AI accelerator, which directly competes with NVIDIAs H100 and has been adopted by major cloud providers, threatening NVIDIAs market share in the data center segment.retrieved_docs_for_baseline_adv baseline_retriever.invoke(complex_query_adv)baseline_contexts [[doc.page_content for doc in retrieved_docs_for_baseline_adv]]advanced_contexts_flat []for step in final_state[past_steps]: advanced_contexts_flat.extend([doc.page_content for doc in step[retrieved_docs]])advanced_contexts [list(set(advanced_contexts_flat))]eval_data { question: [complex_query_adv, complex_query_adv], answer: [baseline_result, final_state[final_answer]], contexts: baseline_contexts advanced_contexts, ground_truth: [ground_truth_answer_adv, ground_truth_answer_adv]}eval_dataset Dataset.from_dict(eval_data)metrics [ context_precision, context_recall, faithfulness, answer_correctness,]print(Running RAGAs evaluation...)result evaluate(eval_dataset, metricsmetrics, is_asyncFalse)print(Evaluation complete.)results_df result.to_pandas()results_df.index [baseline_rag, deep_thinking_rag]print(\n--- RAGAs Evaluation Results ---)print(results_df[[context_precision, context_recall, faithfulness, answer_correctness]].T)输出示例#### OUTPUT ####Preparing evaluation dataset...Running RAGAs evaluation...Evaluation complete.--- RAGAs Evaluation Results --- baseline_rag deep_thinking_ragcontext_precision 0.500000 0.890000context_recall 0.333333 1.000000faithfulness 1.000000 1.000000answer_correctness 0.395112 0.991458量化结果为 Deep Thinking 架构给出明确客观的优势Context Precision0.50 vs 0.89baseline 只有一半相关因为只能检回关于“竞争”的泛化信息advanced agent 通过多步骤、多工具检索显著提升精度。Context Recall0.33 vs 1.00baseline 完全错过了 web 信息召回低advanced 通过规划与工具使用找齐全部必要信息达到满分。Faithfulness1.00 vs 1.00两者都很忠实。baseline 正确指出自己没有信息advanced 正确使用了找到的信息。忠实但不正确的答案也没意义。Answer Correctness0.40 vs 0.99最终质量指标。baseline 因缺失第二部分分析而低于 40%advanced 接近完美。总结整个流水线本文中我们从一个简单、脆弱的 RAG 流水线构建到一个复杂的自治推理 agent先搭建 vanilla RAG并演示它在复杂多源查询上的必然失败系统化地打造 Deep Thinking Agent赋予其规划、多工具使用、与自适应检索策略的能力构建多阶段检索漏斗先广召回hybrid search再高精度cross-encoder reranker最后综合distiller agent使用 LangGraph 编排整个认知架构创建循环、有状态的工作流实现真正的多步推理加入自我批判循环让 agent 能识别失败、修订计划、并在无法得到答案时优雅退出最后用 RAGAs 做生产级评估客观量化证明 advanced agent 的优越性能。使用马尔可夫决策过程MDP学习策略目前我们的 Policy Agent决定CONTINUE或FINISH依赖于像 GPT-4o 这样的通用 LLM每次都要调用。尽管有效但在生产环境可能较慢且昂贵。学术前沿提出了更优的路径。将 RAG 建模为 Decision Process把 agent 的推理循环建模为 Markov Decision ProcessMDP。在这个模型中每个RAGState是“状态”每个 actionCONTINUE、REVISE、FINISH会把系统带入新状态并获得一定“奖励”比如找到正确答案。从经验中学习我们在 LangSmith 中记录的成千上万次成功/失败的推理轨迹都是宝贵的训练数据。每条轨迹都是 agent 在这个 MDP 中的一个例子。训练 Policy Model利用这些数据可以用 Reinforcement Learning 训练一个更小、更专门的 policy 模型。目标速度与效率。把像 GPT-4o 这样复杂模型的决策能力蒸馏到一个更小例如 7B的模型里使CONTINUE/FINISH的决策更快、更省同时高度针对我们的领域。这是诸多研究如 DeepRAG的核心思想也是自治 RAG 系统优化的下一阶段。如何学习AI大模型我在一线互联网企业工作十余年里指导过不少同行后辈。帮助很多人得到了学习和成长。我意识到有很多经验和知识值得分享给大家也可以通过我们的能力和经验解答大家在人工智能学习中的很多困惑所以在工作繁忙的情况下还是坚持各种整理和分享。但苦于知识传播途径有限很多互联网行业朋友无法获得正确的资料得到学习提升故此将并将重要的AI大模型资料包括AI大模型入门学习思维导图、精品AI大模型学习书籍手册、视频教程、实战学习等录播视频免费分享出来。这份完整版的大模型 AI 学习和面试资料已经上传CSDN朋友们如果需要可以微信扫描下方CSDN官方认证二维码免费领取【保证100%免费】第一阶段从大模型系统设计入手讲解大模型的主要方法第二阶段在通过大模型提示词工程从Prompts角度入手更好发挥模型的作用第三阶段大模型平台应用开发借助阿里云PAI平台构建电商领域虚拟试衣系统第四阶段大模型知识库应用开发以LangChain框架为例构建物流行业咨询智能问答系统第五阶段大模型微调开发借助以大健康、新零售、新媒体领域构建适合当前领域大模型第六阶段以SD多模态大模型为主搭建了文生图小程序案例第七阶段以大模型平台应用与开发为主通过星火大模型文心大模型等成熟大模型构建大模型行业应用。学会后的收获• 基于大模型全栈工程实现前端、后端、产品经理、设计、数据分析等通过这门课可获得不同能力• 能够利用大模型解决相关实际项目需求 大数据时代越来越多的企业和机构需要处理海量数据利用大模型技术可以更好地处理这些数据提高数据分析和决策的准确性。因此掌握大模型应用开发技能可以让程序员更好地应对实际项目需求• 基于大模型和企业数据AI应用开发实现大模型理论、掌握GPU算力、硬件、LangChain开发框架和项目实战技能 学会Fine-tuning垂直训练大模型数据准备、数据蒸馏、大模型部署一站式掌握• 能够完成时下热门大模型垂直领域模型训练能力提高程序员的编码能力 大模型应用开发需要掌握机器学习算法、深度学习框架等技术这些技术的掌握可以提高程序员的编码能力和分析能力让程序员更加熟练地编写高质量的代码。1.AI大模型学习路线图2.100套AI大模型商业化落地方案3.100集大模型视频教程4.200本大模型PDF书籍5.LLM面试题合集6.AI产品经理资源合集获取方式有需要的小伙伴可以保存图片到wx扫描二v码免费领取【保证100%免费】