AI创想

标题: LangGraph快速搭建新一代信息检索增强工具DeepResearch [打印本页]

---

作者: 米落枫    时间: 昨天 23:16
标题: LangGraph快速搭建新一代信息检索增强工具DeepResearch
作者：CSDN博客
LangGraph快速搭建新一代信息检索增强工具DeepResearch

项目概览

本项目是一个基于LangGraph框架构建的自主研究智能体，能够根据用户查询自动生成深度研究报告。通过三阶段流水线架构，实现从问题理解到报告生成的完整闭环。

(, 下载次数: 0)

上传

点击文件名下载附件

项目结构

1. langgpraph_deepresearch/
2. ├── .env                           # 环境变量配置
3. ├── langgraph.json                # LangGraph配置文件
4. ├── requirements.txt              # 项目依赖
5. └── graph.py                      # 核心图结构实现

复制代码

环境配置

依赖清单（requirements.txt）

1. langgraph
2. langchain-core
3. langchain-deepseek
4. python-dotenv
5. langsmith
6. pydantic
7. matplotlib
8. seaborn
9. pandas
10. IPython
11. langchain_mcdapters
12. uv

复制代码

环境变量（.env）

1. DEEPSEEK_API_KEY ='****'
2. TAVILY_API_KEY ='tvly-dev-******'
3. LANGSMITH_API_KEY ='lsv2_pt_**********'
4. LANGSMITH_TRACING =trueLANGSMITH_PROJECT='langgpraph_deepresearch'

复制代码

LangGraph配置（langgraph.json）

1. {"dependencies":["./"],"graphs":{"langgpraph_deepresearch":"./graph.py:graph"},"env":".env"}

复制代码

核心代码架构

1. 数据模型定义

1. classWebSearchItem(BaseModel):
2.     query:str
3.     reason:strclassWebSearchPlan(BaseModel):
4.     searches: List[WebSearchItem]classReportData(BaseModel):
5.     short_summary:str
6.     markdown_report:str
7.     follow_up_questions: List[str]

复制代码

2. 规划器（Planner）

1. PLANNER_INSTRUCTIONS =("You are a helpful research assistant. Given a query, come up with 5-7 web searches ""to perform to best answer the query.""Return **ONLY valid JSON** that follows this schema:"'{"searches": [ {"query": "example", "reason": "why"} ]}')
3. planner_chain =(
4.     planner_prompt
5.     | model.with_structured_output(WebSearchPlan, method="json_mode"))

复制代码

3. 搜索代理（Search Agent）

1. SEARCH_INSTRUCTIONS =("You are a research assistant. Given a search term, you search the web for that term and ""produce a concise summary of the results. The summary must 2-3 paragraphs and less than 300 ""words. Capture the main points. Write succinctly, no need to have complete sentences or good ""grammar. This will be consumed by someone synthesizing a report, so its vital you capture the ""essence and ignore any fluff. Do not include any additional commentary other than the summary ""itself.")
3. search_tool = TavilySearch(max_results=5, topic="general")
4. search_agent = create_react_agent(
5.     model=model,
6.     prompt=SEARCH_INSTRUCTIONS,
7.     tools=[search_tool],)

复制代码

4. 写作器（Writer）

1. WRITER_PROMPT =("You are a senior researcher tasked with writing a cohesive report for a research query. ""You will be provided with the original query and some initial research.\n\n""① 先给出完整的大纲。\n""② 然后生成正式报告。\n""**写作要求**：\n""· 报告使用 Markdown 格式；\n""· 章节清晰，层次分明；\n""· markdown_report部分至少包含2000中文字符（注意需要用中文进行回复）；\n""· 内容丰富，论据充分，可加入引用和数据，允许分段、添加引用、表格等；\n""· 最终仅返回 JSON：\n"'{"short_summary": "...", "markdown_report": "...", "follow_up_questions": ["..."]}')
3. writer_chain = writer_prompt | model.with_structured_output(ReportData, method="json_mode")

复制代码

LangGraph节点实现

1. 规划节点（planner_node）

1. defplanner_node(state: MessagesState)-> Command:
2.     user_query = state["messages"][-1].content
3.     raw = planner_chain.invoke({"query": user_query})print(raw)try:
4.         plan = WebSearchPlan.model_validate(raw)except ValidationError:ifisinstance(raw,dict)andisinstance(raw.get("searches"),list):
5.             plan = WebSearchPlan(searches=[WebSearchItem(query=q, reason="")for q in raw["searches"]])else:raisereturn Command(goto="search_node", update={"messages":[AIMessage(content=plan.model_dump_json())],"plan": plan})

复制代码

2. 搜索节点（search_node）

1. defsearch_node(state: MessagesState)-> Command:
2.     plan_json = state["messages"][-1].content
3.     plan = WebSearchPlan.model_validate_json(plan_json)
4.     summaries =[]for item in plan.searches:# 串行处理
5.         run = search_agent.invoke({"messages":[HumanMessage(content=item.query)]})
6.         msgs = run["messages"]
7.         readable =next((m for m inreversed(msgs)ifisinstance(m,(ToolMessage, AIMessage))), msgs[-1])
8.         summaries.append(f"##{item.query}\n\n{readable.content}")
9.    
10.     combined ="\n\n".join(summaries)return Command(goto="writer_node", update={"messages":[AIMessage(content=combined)]})

复制代码

3. 写作节点（writer_node）

1. defwriter_node(state: MessagesState)-> Command:
2.     original_query = state["messages"][0].content
3.     combined_summary = state["messages"][-1].content
4.    
5.     writer_input =(f"原始问题：{original_query}\n\n"f"搜索摘要：\n{combined_summary}")
6.    
7.     report = writer_chain.invoke({"content": writer_input})return Command(
8.         goto=END,
9.         update={"messages":[AIMessage(content=json.dumps(report.dict(), ensure_ascii=False, indent=4))]})

复制代码

图构建与执行

1. builder = StateGraph(MessagesState)
2. builder.add_node("planner_node", planner_node)
3. builder.add_node("search_node", search_node)
4. builder.add_node("writer_node", writer_node)
6. builder.add_edge(START,"planner_node")
7. builder.add_edge("planner_node","search_node")
8. builder.add_edge("search_node","writer_node")
9. builder.add_edge("writer_node", END)
11. graph = builder.compile()

复制代码

运行流程

输入：用户查询通过messages参数传入规划：LLM生成5-7个搜索计划搜索：串行执行每个搜索任务综合：整合所有搜索结果写作：生成结构化中文报告输出：返回包含摘要、正文、后续问题的JSON

关键特性

中文支持：全程中文交互和报告生成结构化输出：使用Pydantic保证数据格式容错机制：ValidationError的优雅处理LangSmith集成：完整的调用链追踪可扩展性：模块化设计便于功能扩展

性能分析

串行搜索：当前为顺序执行，适合稳定性优先的场景Token优化：每次搜索限制在300词以内缓存支持：LangGraph自动状态管理调试友好：print语句和LangSmith双重调试

完整的项目代码graph.py

1. # 封装更完整的图结构import json
2. import os
3. from typing import List
5. from dotenv import load_dotenv
6. from pydantic import BaseModel, ValidationError, parse_obj_as
8. from langchain_deepseek import ChatDeepSeek
9. from langchain.prompts import ChatPromptTemplate
10. from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
11. from langgraph.graph import StateGraph, MessagesState, START, END
12. from langgraph.types import Command
13. from langgraph.prebuilt import create_react_agent
14. from langchain_tavily import TavilySearch
15. from langchain_openai import ChatOpenAI
17. load_dotenv()
18. model = ChatDeepSeek(model="deepseek-chat", max_tokens=8000)# model = ChatOpenAI(model="gpt-4-1", max_tokens=32000)# -------- 1) Planner Chain --------
19. PLANNER_INSTRUCTIONS =("You are a helpful research assistant. Given a query, come up with 5-7 web searches ""to perform to best answer the query.""Return **ONLY valid JSON** that follows this schema:"'{{"searches": [ {{"query": "example", "reason": "why"}} ]}}')classWebSearchItem(BaseModel):
20.     query:str
21.     reason:strclassWebSearchPlan(BaseModel):
22.     searches: List[WebSearchItem]
24. planner_prompt = ChatPromptTemplate.from_messages([("system", PLANNER_INSTRUCTIONS),("human","{query}")])
26. planner_chain =(
27.     planner_prompt
28.     | model.with_structured_output(WebSearchPlan, method="json_mode")# 强制 JSON)# -------- 2) search agent --------
29. SEARCH_INSTRUCTIONS =("You are a research assistant. Given a search term, you search the web for that term and ""produce a concise summary of the results. The summary must 2-3 paragraphs and less than 300 ""words. Capture the main points. Write succinctly, no need to have complete sentences or good ""grammar. This will be consumed by someone synthesizing a report, so its vital you capture the ""essence and ignore any fluff. Do not include any additional commentary other than the summary ""itself.")
31. search_tool = TavilySearch(max_results=5, topic="general")
33. search_agent = create_react_agent(
34.     model=model,
35.     prompt=SEARCH_INSTRUCTIONS,
36.     tools=[search_tool],)# -------- 3) Writer Chain --------
37. WRITER_PROMPT =("You are a senior researcher tasked with writing a cohesive report for a research query. ""You will be provided with the original query and some initial research.\n\n""① 先给出完整的大纲。\n""② 然后生成正式报告。\n""**写作要求**：\n""· 报告使用 Markdown 格式；\n""· 章节清晰，层次分明；\n""· markdown_report部分至少包含2000中文字符（注意需要用中文进行回复）；\n""· 内容丰富，论据充分，可加入引用和数据，允许分段、添加引用、表格等；\n""· 最终仅返回 JSON：\n"'{{"short_summary": "...", "markdown_report": "...", "follow_up_questions": ["..."]}}')classReportData(BaseModel):
38.     short_summary:str
39.     markdown_report:str
40.     follow_up_questions: List[str]
42. writer_prompt = ChatPromptTemplate.from_messages([("system", WRITER_PROMPT),("human","{content}")])
43. writer_chain = writer_prompt | model.with_structured_output(ReportData, method="json_mode")# -------------LangGraph 节点----------------defplanner_node(state: MessagesState)-> Command:
44.     user_query = state["messages"][-1].content
45.     raw = planner_chain.invoke({"query": user_query})print(raw)try:# plan = parse_obj_as(WebSearchPlan, raw)
46.         plan = WebSearchPlan.model_validate(raw)except ValidationError:ifisinstance(raw,dict)andisinstance(raw.get("searches"),list):
47.             plan = WebSearchPlan(searches=[WebSearchItem(query=q, reason="")for q in raw["searches"]])else:raisereturn Command(goto="search_node", update={"messages":[AIMessage(content=plan.model_dump_json())],"plan": plan})# -------------search_node----------------defsearch_node(state: MessagesState)-> Command:
48.     plan_json = state["messages"][-1].content
49.     plan = WebSearchPlan.model_validate_json(plan_json)
50.     summaries =[]for item in plan.searches:
51.         run = search_agent.invoke({"messages":[HumanMessage(content=item.query)]})
52.         msgs = run["messages"]
53.         readable =next((m for m inreversed(msgs)ifisinstance(m,(ToolMessage, AIMessage))), msgs[-1])
54.         summaries.append(f"##{item.query}\n\n{readable.content}")
55.     combined ="\n\n".join(summaries)return Command(goto="writer_node", update={"messages":[AIMessage(content=combined)]})# -------------write_node----------------defwriter_node(state: MessagesState)-> Command:
56.     original_query = state["messages"][0].content
57.     combined_summary = state["messages"][-1].content
58.    
59.     writer_input =(f"原始问题：{original_query}\n\n"f"搜索摘要：\n{combined_summary}")
60.    
61.     report = writer_chain.invoke({"content": writer_input})return Command(
62.         goto=END,
63.         update={"messages":[AIMessage(content=json.dumps(report.dict(), ensure_ascii=False, indent=4))]})# 构建并运行Graph
64. builder = StateGraph(MessagesState)
65. builder.add_node("planner_node", planner_node)
66. builder.add_node("search_node", search_node)
67. builder.add_node("writer_node", writer_node)# 定义节点间的连接关系
68. builder.add_edge(START,"planner_node")
69. builder.add_edge("planner_node","search_node")
70. builder.add_edge("search_node","writer_node")
71. builder.add_edge("writer_node", END)# 编译Graph
72. graph = builder.compile()

复制代码

使用示例

1. # 运行研究
2. result = graph.invoke({"messages":[HumanMessage(content="分析2024年人工智能发展趋势")]})# 解析结果import json
3. report_data = json.loads(result["messages"][-1].content)print("摘要：", report_data["short_summary"])print("报告：", report_data["markdown_report"])

复制代码

部署

项目部署大家参考这里：langgraph-cli快速完成项目部署

(, 下载次数: 0)

上传

点击文件名下载附件

扩展方向

并行化搜索：使用asyncio.gather优化性能缓存机制：避免重复搜索相同主题多语言支持：扩展至英文等多语言报告自定义模板：支持不同领域的报告模板

总结

DeepResearch项目展示了如何基于LangGraph构建一个完整的AI研究智能体，通过三阶段流水线实现了从问题理解到深度报告生成的完整闭环。项目代码结构清晰，配置完整，具备良好的可扩展性和实用性。

---

参考：赋范空间大模型技术社区–DeepResearch应用开发实战

原文地址：https://blog.csdn.net/Galen_xia/article/details/149817138

---

欢迎光临 AI创想 (https://www.llms-ai.com/)

Powered by Discuz! X3.4