167 lines
6.2 KiB
Python
167 lines
6.2 KiB
Python
import os
|
|
from config import OPENAI_API_KEY,LANGCHAIN_API_KEY
|
|
from langchain_openai import ChatOpenAI
|
|
from langchain.tools import tool
|
|
from langchain_core.messages import HumanMessage, SystemMessage
|
|
from langgraph.graph import StateGraph, START, END
|
|
from langgraph.prebuilt import ToolNode
|
|
from typing import TypedDict # 定义数据类型
|
|
from typing import Annotated # 注释说明细节
|
|
from langgraph.graph.message import add_messages
|
|
|
|
# LangSmith调试
|
|
os.environ["LANGCHAIN_TRACING_V2"] = "true" # 总开关，决定启用追踪功能
|
|
os.environ["LANGCHAIN_PROJECT"] = "multi_agent_orchestration" # 自定义项目名
|
|
os.environ["LANGCHAIN_API_KEY"] = LANGCHAIN_API_KEY
|
|
|
|
# llm配置
|
|
llm = ChatOpenAI(
|
|
model="deepseek-chat",
|
|
api_key=OPENAI_API_KEY,
|
|
base_url="https://api.deepseek.com"
|
|
)
|
|
|
|
# 模拟工具
|
|
@tool
|
|
def search_internal_docs(query:str):
|
|
"""搜索公司内部文档获取政策信息"""
|
|
return "根据公司手册，年假为15天"
|
|
|
|
@tool
|
|
def search_web(query:str):
|
|
"""通过搜索引擎获取最新的公开信息"""
|
|
return "据TechCrunch报道，LangGraph 0.6已支持持久化记忆"
|
|
|
|
@tool
|
|
def generate_code(requirement:str):
|
|
"""根据需求生成可运行的Python代码"""
|
|
return "python\nprint('Hello from Code writer!')"
|
|
|
|
|
|
# 共享状态定义
|
|
class AgentState(TypedDict):
|
|
messages: Annotated[list,add_messages] # 自动累积对话历史
|
|
next_speaker: str
|
|
|
|
# 专家节点
|
|
def rag_expert(state:AgentState):
|
|
prompt = "你是公司知识库专家，只基于内部文档回答问题。回答应简洁明了，直接给出最终结论。请在回答的最后一行加上：(任务已完成)"
|
|
messages = [SystemMessage(content=prompt)]+state['messages']
|
|
tools = [search_internal_docs]
|
|
response = llm.bind_tools(tools).invoke(messages)
|
|
return {'messages':[response]}
|
|
|
|
def web_research(state:AgentState):
|
|
prompt = "你是互联网研究员，擅长用搜索引擎获取最新公开信息。回答应简洁明了，直接给出最终结论。请在回答的最后一行加上：(任务已完成)"
|
|
messages = [SystemMessage(content=prompt)]+state['messages']
|
|
tools = [search_web]
|
|
response = llm.bind_tools(tools).invoke(messages)
|
|
return {'messages':[response]}
|
|
|
|
def code_writer(state:AgentState):
|
|
prompt = "你是python工程师，只生成可运行代码，不解释。回答应简洁明了，直接给出最终结论。请在回答的最后一行加上：(任务已完成)"
|
|
messages = [SystemMessage(content=prompt)]+state['messages']
|
|
tools = [generate_code]
|
|
response = llm.bind_tools(tools).invoke(messages)
|
|
return {'messages':[response]}
|
|
|
|
# 总控节点
|
|
def supervisor(state:AgentState):
|
|
supervisor_prompt = """
|
|
你是一个任务协调员。你的目标是管理专家来解决用户的问题。
|
|
|
|
当前对话需要以下专家参与：
|
|
- rag_expert：涉及公司政策、内部流程
|
|
- web_research：涉及外部新闻、公开数据
|
|
- code_writer：需要生成代码
|
|
|
|
【决策逻辑】
|
|
1. **检查历史记录**：先看上一个回复是否已经完整回答了用户的初始问题。
|
|
2. **如果已经回答完毕**：必须输出 'FINISH'。
|
|
3. **如果尚未回答或需要补充**：根据当前缺少的步骤，选择下一个最合适的专家。
|
|
|
|
请只输出专家名字或 'FINISH'，不要输出任何其他解释。
|
|
"""
|
|
messages = [SystemMessage(content=supervisor_prompt)]+state['messages']
|
|
response = llm.invoke(messages)
|
|
next_speaker = response.content.strip()
|
|
return {"next_speaker":next_speaker}
|
|
|
|
# 路由函数定义
|
|
def route_supervisor(state:AgentState):
|
|
if state["next_speaker"]=="FINISH":
|
|
return END
|
|
return state["next_speaker"]
|
|
|
|
def should_continue(state:AgentState):
|
|
last_msg = state["messages"][-1]
|
|
if hasattr(last_msg,"tool_calls") and last_msg.tool_calls:
|
|
return "tools"
|
|
return "supervisor"
|
|
|
|
def route_after_tool(state:AgentState):
|
|
# 工具执行完后，通过next_speaker知道是谁调用的，路由回去
|
|
return state["next_speaker"]
|
|
|
|
|
|
# 添加工具节点
|
|
tools = [search_internal_docs,search_web,generate_code]
|
|
tool_node = ToolNode(tools)
|
|
|
|
|
|
# 构建协作图
|
|
workflow = StateGraph(AgentState)
|
|
|
|
# 1. 添加节点
|
|
workflow.add_node("supervisor",supervisor)
|
|
workflow.add_node("rag_expert",rag_expert)
|
|
workflow.add_node("web_research",web_research)
|
|
workflow.add_node("code_writer",code_writer)
|
|
workflow.add_node("tools",tool_node)
|
|
|
|
|
|
# 2. 总控回路
|
|
workflow.add_edge(START,"supervisor")
|
|
workflow.add_conditional_edges("supervisor",route_supervisor)
|
|
|
|
# 3. 专家节点的ReAct循环
|
|
for member in ["rag_expert","web_research","code_writer"]: # 为每个专家添加条件边：决定是去执行工具还是回总控
|
|
workflow.add_conditional_edges(
|
|
member,
|
|
should_continue,
|
|
{"tools":"tools","supervisor":"supervisor"}
|
|
)
|
|
|
|
# 4.工具节点闭环
|
|
workflow.add_conditional_edges( # 工具执行完，根据next_speaker路由回原来的专家
|
|
"tools",
|
|
route_after_tool
|
|
)
|
|
|
|
app = workflow.compile()
|
|
|
|
|
|
# 测试运行
|
|
if __name__ == '__main__':
|
|
user_input = "公司年假多少天"
|
|
print("用户提问:",user_input)
|
|
print('\n开始多智能体协作...\n')
|
|
|
|
inputs = {"messages":[HumanMessage(content=user_input)]}
|
|
# app是编译好的图，stream()会让图开始运转，并返回一个生成器
|
|
# 图每执行完一个节点，就会产出一个step字典
|
|
for step in app.stream(inputs):
|
|
# 因为step是个字典，所以需要拆包拿到 节点名(Node) 与 输出内容(output)
|
|
for node,output in step.items():
|
|
# 有工具/专家回复
|
|
if "messages" in output:
|
|
msg = output["messages"][-1]
|
|
if hasattr(msg,"tool_calls") and msg.tool_calls:
|
|
call = msg.tool_calls[0]
|
|
print(f"【{node}】调用工具 {call['name']}({call['args']})")
|
|
else:
|
|
print(f"【{node}】回复：{msg.content}")
|
|
# supervisor刚做完决策，确定下个发言人
|
|
elif "next_speaker" in output:
|
|
speaker = output["next_speaker"]
|
|
print(f"【Supervisor】指定下一位发言人：{speaker}") |