feat(langgraph): introduce advanced agent patterns in new module
This commit is contained in:
Annyfee
@@ -0,0 +1,118 @@
|
||||
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
|
||||
from langgraph.graph import StateGraph, MessagesState, START, END
|
||||
from langgraph.prebuilt import ToolNode
|
||||
from langgraph.checkpoint.memory import MemorySaver
|
||||
|
||||
# LangSmith调试
|
||||
os.environ["LANGCHAIN_TRACING_V2"] = "true" # 总开关，决定启用追踪功能
|
||||
os.environ["LANGCHAIN_PROJECT"] = "human_approval" # 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"
|
||||
)
|
||||
|
||||
|
||||
# 定义一个敏感工具：发送邮件(模拟)
|
||||
@tool
|
||||
def send_email(to, content):
|
||||
"""模拟发送邮件"""
|
||||
return f'邮件已发送至{to},内容为：{content}'
|
||||
|
||||
|
||||
# 工具绑定到llm
|
||||
tools = [send_email]
|
||||
llm_with_tools = llm.bind_tools(tools)
|
||||
|
||||
# Node函数与Edge节点
|
||||
tool_node = ToolNode(tools)
|
||||
|
||||
|
||||
def call_model(state: MessagesState):
|
||||
response = llm_with_tools.invoke(state['messages'])
|
||||
return {"messages":[response]}
|
||||
|
||||
|
||||
def should_continue(state: MessagesState):
|
||||
last_msg = state['messages'][-1]
|
||||
if hasattr(last_msg, "tool_calls") and last_msg.tool_calls:
|
||||
return "tools"
|
||||
return END
|
||||
|
||||
|
||||
|
||||
# 构建基础ReAct图
|
||||
workflow = StateGraph(MessagesState)
|
||||
|
||||
workflow.add_node("agent", call_model)
|
||||
workflow.add_node("tools", tool_node)
|
||||
workflow.add_edge(START, "agent")
|
||||
|
||||
workflow.add_conditional_edges(
|
||||
"agent",
|
||||
should_continue,
|
||||
{
|
||||
"tools": "tools",
|
||||
END: END
|
||||
}
|
||||
)
|
||||
|
||||
workflow.add_edge("tools", "agent")
|
||||
|
||||
app = workflow.compile(
|
||||
# 在内存里做状态持久化
|
||||
checkpointer=MemorySaver(),
|
||||
interrupt_before=["tools"] # 选择要人工审批的节点 -- 负责在哪里停，之后的代码负责停了之后怎么办
|
||||
)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
config = {
|
||||
"configurable":{"thread_id":"user123"}
|
||||
}
|
||||
user_input = "请帮我给 boss@example.com 发一封邮件，内容是：会议推迟到明天下午3点。"
|
||||
|
||||
print("用户输入:",user_input)
|
||||
print("\nAgent正在思考...\n")
|
||||
|
||||
# 初识输入
|
||||
inputs = {"messages":[HumanMessage(content=user_input)]}
|
||||
|
||||
while 1:
|
||||
# 触发工作流执行，推进到下一个中断点或自然结束
|
||||
# inputs注入事件;config确定回话id;"values":完整记录每步结果
|
||||
for _ in app.stream(inputs,config,stream_mode="values"): # 流式执行
|
||||
pass # 必须迭代生成器，才能实际执行工作流
|
||||
|
||||
# 获取当前状态
|
||||
snapshot = app.get_state(config)
|
||||
next_tasks = snapshot.next # 返回下一步要执行的节点名列表
|
||||
|
||||
# 如果没有下一步，说明工作流已结束
|
||||
if not next_tasks:
|
||||
final_msg = snapshot.values['messages'][-1]
|
||||
print(f'\n最终回复:{final_msg.content}')
|
||||
break
|
||||
|
||||
# 如果下一步是需要审批的节点
|
||||
if "tools" in next_tasks:
|
||||
last_msg = snapshot.values['messages'][-1]
|
||||
tool_call = last_msg.tool_calls[0]
|
||||
print(f'\n⚠️ Agent准备执行操作：')
|
||||
print(f' 工具名称：{tool_call["name"]}')
|
||||
print(f' 参数：{tool_call["args"]}')
|
||||
|
||||
approval = input("\n✅ 是否批准执行？(输入 'yes' 继续，其他取消): ").strip().lower()
|
||||
if approval == "yes":
|
||||
print('\n 继续执行...')
|
||||
inputs = None # 表示从断点继续，无新输入
|
||||
else:
|
||||
print("\n❌ 操作已取消，流程终止")
|
||||
break
|
||||
@@ -0,0 +1,119 @@
|
||||
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, MessagesState, START, END
|
||||
from langgraph.prebuilt import ToolNode
|
||||
from typing import TypedDict
|
||||
|
||||
# LangSmith调试
|
||||
os.environ["LANGCHAIN_TRACING_V2"] = "true" # 总开关，决定启用追踪功能
|
||||
os.environ["LANGCHAIN_PROJECT"] = "graph_as_tool" # 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"
|
||||
)
|
||||
|
||||
# 构建子工作流
|
||||
# 1.子任务状态
|
||||
class RetryState(TypedDict):
|
||||
query: str
|
||||
attempt: int
|
||||
result: str
|
||||
|
||||
# 2.子图逻辑 -- 模拟一个可能失败，需重试的API调用
|
||||
def call_unstable_api(state:RetryState):
|
||||
"""模拟偶发性的外部服务，偶发失败"""
|
||||
attempt = state["attempt"]
|
||||
if attempt == 1:
|
||||
# 第一次故意失败
|
||||
return {"result":"ERROR：服务暂时不可用","attempt":attempt+1}
|
||||
else:
|
||||
# 第二次成功
|
||||
return {"result":f"SUCCESS：成功处理请求：{state['query']}","attempt":attempt+1}
|
||||
|
||||
def should_retry(state:RetryState):
|
||||
if "ERROR" in state["result"] and state["attempt"] <= 2: # 出现报错且重试次数小于2，重连
|
||||
return "call_api"
|
||||
return END
|
||||
|
||||
# 3.构建子图工作流
|
||||
retry_workflow = StateGraph(RetryState)
|
||||
retry_workflow.add_node("call_api",call_unstable_api)
|
||||
retry_workflow.add_edge(START,"call_api")
|
||||
retry_workflow.add_conditional_edges(
|
||||
"call_api",
|
||||
should_retry,
|
||||
{"call_api":"call_api",END:END}
|
||||
)
|
||||
retry_app = retry_workflow.compile()
|
||||
|
||||
# 4.封装为tool(Graph-as-a-Tool)
|
||||
@tool
|
||||
def create_order(query:str) -> str:
|
||||
"""创建新订单，自动重试保障成功率"""
|
||||
result = retry_app.invoke({"query":query,"attempt":1,"result":""})
|
||||
return result["result"]
|
||||
|
||||
|
||||
# 5.主graph
|
||||
tools = [create_order]
|
||||
llm_with_tools = llm.bind_tools(tools)
|
||||
tool_node = ToolNode(tools)
|
||||
|
||||
def agent_node(state:MessagesState):
|
||||
response = llm_with_tools.invoke(state["messages"])
|
||||
return {"messages":[response]}
|
||||
|
||||
def should_continue(state:MessagesState):
|
||||
last_msg = state["messages"][-1]
|
||||
if hasattr(last_msg,"tool_calls") and last_msg.tool_calls:
|
||||
return "tools"
|
||||
return END
|
||||
|
||||
# 构建主工作流
|
||||
workflow = StateGraph(MessagesState)
|
||||
workflow.add_node("agent",agent_node)
|
||||
workflow.add_node("tools",tool_node)
|
||||
workflow.add_edge(START,"agent")
|
||||
workflow.add_conditional_edges(
|
||||
"agent",
|
||||
should_continue,
|
||||
{
|
||||
"tools": "tools",
|
||||
END: END
|
||||
}
|
||||
)
|
||||
workflow.add_edge("tools","agent")
|
||||
|
||||
app = workflow.compile()
|
||||
|
||||
|
||||
# 运行
|
||||
if __name__ == '__main__':
|
||||
user_input = "请创建一个新订单：购买三本书"
|
||||
print('用户输入:',user_input)
|
||||
|
||||
inputs = {"messages":[
|
||||
SystemMessage(content="你是一个任务执行助手。当用户提出任何需要处理、操作或执行的请求时，必须调用 create_order 工具来完成，不要自行回答细节"),
|
||||
HumanMessage(content=user_input)
|
||||
]}
|
||||
result = app.invoke(inputs)
|
||||
|
||||
tool_result = None
|
||||
# 在主工作流的消息历史中，查找最近的工具执行结果
|
||||
for msg in reversed(result["messages"]):
|
||||
if msg.type == "tool": # 找到ToolMessage类型消息
|
||||
tool_result = msg.content
|
||||
break
|
||||
if tool_result:
|
||||
print(f"\n✅ 直接获取子图返回值:\n{tool_result}")
|
||||
else:
|
||||
print("\n❌ 未执行任何工具")
|
||||
final_reply = result["messages"][-1]
|
||||
print(f'\n最终回复:\n{final_reply}')
|
||||
@@ -0,0 +1,167 @@
|
||||
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, MessagesState, 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}")
|
||||
@@ -0,0 +1,227 @@
|
||||
import os
|
||||
import time
|
||||
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,MessagesState,START,END
|
||||
from langgraph.prebuilt import ToolNode
|
||||
from langgraph.checkpoint.memory import MemorySaver
|
||||
from typing import TypedDict,Annotated
|
||||
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"] = "supervisor_agent_ops_system" # 自定义项目名
|
||||
os.environ["LANGCHAIN_API_KEY"] = LANGCHAIN_API_KEY
|
||||
|
||||
llm = ChatOpenAI(
|
||||
model="deepseek-chat",
|
||||
api_key=OPENAI_API_KEY,
|
||||
base_url="https://api.deepseek.com"
|
||||
)
|
||||
|
||||
# === 一、Graph-as-a-Tool ===
|
||||
# === 模拟一个不稳定的SSH日志查询过程 ===
|
||||
|
||||
class SSHState(TypedDict):
|
||||
target_ip: str
|
||||
attempt: int
|
||||
logs: str
|
||||
|
||||
def connect_ssh(state:SSHState):
|
||||
"""模拟SSH连接，第一次连接必定超时"""
|
||||
print(f' [子图]正在尝试连接服务器{state["target_ip"]}(第{state["attempt"]}次)')
|
||||
if state['attempt'] == 1:
|
||||
return {'logs':'ERROR:Connection Timed Out','attempt':state['attempt']+1}
|
||||
return {'logs':'CONNECTED','attempt':state['attempt']+1}
|
||||
|
||||
def grep_system_logs(state:SSHState):
|
||||
"""连接成功后读取日志"""
|
||||
if state["logs"] == "CONNECTED":
|
||||
# 打印查到的结果
|
||||
return {'logs':f'SUCCESS: Retrieved logs from {state['target_ip']}:[ERROR: OutOfMemory at line 4032]'}
|
||||
return {'logs':state['logs']} # 保持错误状态
|
||||
|
||||
def ssh_routing(state:SSHState):
|
||||
"""路由逻辑：如果连接失败且尝试次数少于3，重试"""
|
||||
if "ERROR" in state['logs'] and state["attempt"] <= 2:
|
||||
return "connect"
|
||||
return "grep"
|
||||
|
||||
# 构建子图
|
||||
ssh_workflow = StateGraph(SSHState)
|
||||
ssh_workflow.add_node("connect",connect_ssh)
|
||||
ssh_workflow.add_node("grep",grep_system_logs)
|
||||
|
||||
ssh_workflow.add_edge(START,"connect")
|
||||
ssh_workflow.add_conditional_edges("connect",ssh_routing,{"connect":"connect","grep":"grep"})
|
||||
ssh_workflow.add_edge("grep",END)
|
||||
|
||||
ssh_app = ssh_workflow.compile()
|
||||
|
||||
# 将子图封装为工具
|
||||
@tool
|
||||
def analyze_server_logs(ip_address:str):
|
||||
"""使用SSH连接服务器并分析最近的错误日志（内含自动重连机制）"""
|
||||
result = ssh_app.invoke({"target_ip":ip_address,"attempt":1,"log":""})
|
||||
return result['logs']
|
||||
|
||||
|
||||
|
||||
# === 二、Human-in-the-Loop ===
|
||||
# === 重启服务，高危操作，需要审批 ===
|
||||
|
||||
@tool
|
||||
def restart_service(service_name:str):
|
||||
"""重启指定的服务器服务"""
|
||||
return f"服务[{service_name}]已成功重启，系统负载已恢复正常"
|
||||
|
||||
|
||||
|
||||
# === 三、Multi-Agent 编排 ===
|
||||
# === 总控调度 + 专家分工 ===
|
||||
|
||||
# 1. 共享状态
|
||||
class AgentState(TypedDict):
|
||||
messages:Annotated[list,add_messages]
|
||||
next_speaker:str
|
||||
|
||||
# 2. 专家节点
|
||||
def log_expert(state:AgentState):
|
||||
prompt = "你是日志分析专家，使用工具分析服务器日志，找出报错原因。回答需简洁。"
|
||||
messages = [SystemMessage(content=prompt)] + state['messages']
|
||||
# 绑定子图工具
|
||||
tools = [analyze_server_logs]
|
||||
response = llm.bind_tools(tools).invoke(messages)
|
||||
return {"messages":[response]}
|
||||
|
||||
def ops_expert(state:AgentState):
|
||||
prompt = "你是运维专家。当收到修复指令时，请立即调用 'restart_service' 工具进行修复，不要输出任何额外的解释文本。"
|
||||
messages = [SystemMessage(content=prompt)] + state['messages']
|
||||
tools = [restart_service]
|
||||
# 绑定敏感工具
|
||||
response = llm.bind_tools(tools).invoke(messages)
|
||||
return {"messages":[response]}
|
||||
|
||||
# 3. 总控节点（supervisor）
|
||||
def supervisor(state:AgentState):
|
||||
prompt = """
|
||||
你是 IT 运维总指挥。
|
||||
专家列表：
|
||||
- log_expert
|
||||
- ops_expert
|
||||
|
||||
决策逻辑：
|
||||
1. 未知原因 -> log_expert
|
||||
2. 已知原因（如OOM、报错） -> ops_expert
|
||||
3. 修复完成 -> FINISH
|
||||
|
||||
【输出约束】
|
||||
仅输出下一个专家的名字（如 log_expert），不要包含任何其他字符或标点。
|
||||
"""
|
||||
messages = [SystemMessage(content=prompt)] + state['messages']
|
||||
response = llm.invoke(messages)
|
||||
return {"next_speaker":response.content.strip()}
|
||||
|
||||
# 4. 路由逻辑
|
||||
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):
|
||||
return state["next_speaker"]
|
||||
|
||||
|
||||
# === 四、构建主图与集成 ===
|
||||
|
||||
# 工具集合
|
||||
all_tools = [analyze_server_logs,restart_service]
|
||||
tool_node = ToolNode(all_tools)
|
||||
|
||||
# 构建主图
|
||||
workflow = StateGraph(AgentState)
|
||||
|
||||
workflow.add_node("supervisor",supervisor)
|
||||
workflow.add_node("log_expert",log_expert)
|
||||
workflow.add_node("ops_expert",ops_expert)
|
||||
workflow.add_node("tools",tool_node)
|
||||
|
||||
workflow.add_edge(START,"supervisor")
|
||||
workflow.add_conditional_edges("supervisor",route_supervisor)
|
||||
for member in ["log_expert","ops_expert"]:
|
||||
workflow.add_conditional_edges(
|
||||
member,
|
||||
should_continue,
|
||||
{"tools":"tools","supervisor":"supervisor"}
|
||||
)
|
||||
workflow.add_conditional_edges("tools",route_after_tool)
|
||||
|
||||
# 编译图:加入记忆与中断机制(Human-in-the-Loop)
|
||||
# 注: 我们在所有工具执行前都暂停，但在运行时进行逻辑判断
|
||||
app = workflow.compile(
|
||||
checkpointer=MemorySaver(),
|
||||
interrupt_before=["tools"]
|
||||
)
|
||||
|
||||
# === 五、运行时逻辑(模拟生产环境的交互) ===
|
||||
|
||||
if __name__ == '__main__':
|
||||
# 模拟一次完整的故障处理流程
|
||||
user_input = "服务器 192.168.1.100 报警，响应极慢，请处理。"
|
||||
config = {
|
||||
"configurable": {"thread_id": "incident_001"}
|
||||
}
|
||||
|
||||
print(f'收到报警 : {user_input}')
|
||||
inputs = {"messages":[HumanMessage(content=user_input)]}
|
||||
|
||||
# 循环执行，直到任务结束
|
||||
while 1:
|
||||
# 1. 执行图直到中断或结束
|
||||
for _ in app.stream(inputs,config,stream_mode="values"):
|
||||
pass
|
||||
|
||||
# 2. 检查当前状态
|
||||
snapshot = app.get_state(config)
|
||||
next_tasks = snapshot.next
|
||||
|
||||
# 没有下一步，任务结束
|
||||
if not next_tasks:
|
||||
print(f" 最终报告:{snapshot.values['messages'][-1].content}")
|
||||
break
|
||||
|
||||
# 3. 处理中断:判断是哪个工具被调用
|
||||
if "tools" in next_tasks:
|
||||
last_msg = snapshot.values['messages'][-1]
|
||||
tool_call = last_msg.tool_calls[0]
|
||||
tool_name = tool_call["name"]
|
||||
|
||||
print(f'\n[系统暂停] 请求调用工具:{tool_name}')
|
||||
|
||||
# 策略A: 自动放行安全工具(Graph-as-a-Tool)
|
||||
if tool_name == "analyze_server_logs":
|
||||
print(' -> 这是一个查询类工具，系统自行批准。')
|
||||
inputs = None # 继续执行
|
||||
continue
|
||||
|
||||
# 策略B: 拦截高危工具(Human-in-the-Loop)
|
||||
elif tool_name == "restart_service":
|
||||
print(" -> ⚠️ 警告: 这是一个高危操作!")
|
||||
user_approval = input(" -> 请人工审批 (输入 'yes' 允许重启):")
|
||||
|
||||
if user_approval == "yes":
|
||||
print(' -> ✅ 审批通过，正在执行...')
|
||||
inputs = None # 继续执行
|
||||
else:
|
||||
print(' -> ❌️ 审批拒绝，任务终止!')
|
||||
# 实际系统中，应通过 ToolMessage 反馈人工拒绝，使 LLM 能继续响应；
|
||||
# 当前demo为简化，直接退出
|
||||
break
|
||||
@@ -0,0 +1,102 @@
|
||||
# 🧩 模块说明：LangGraph 进阶篇 - 构建企业级智能 Agent 系统
|
||||
|
||||
> 📌 核心知识点：Human-in-the-Loop 人工干预｜Graph-as-a-Tool 子图封装｜Multi-Agent 多智能体协作｜Supervisor 总控调度｜复杂工作流编排
|
||||
|
||||
---
|
||||
|
||||
### 1. `01_human_approval.py` （人工审批机制）
|
||||
|
||||
实现带有人工干预的智能体工作流，为敏感操作添加审批环节。
|
||||
|
||||
- ✅ 掌握点：
|
||||
- 在 LangGraph 中使用 `interrupt_before=["tools"]` 设置中断点
|
||||
- 通过 `app.get_state(config)` 获取工作流当前状态
|
||||
- 实现用户交互审批流程，控制高危操作的执行
|
||||
- 结合 MemorySaver 实现持久化状态管理
|
||||
|
||||
- 输出：
|
||||
- 当智能体尝试执行敏感操作（如发送邮件）时，系统会暂停并请求用户确认
|
||||
- 用户输入 "yes" 批准执行，输入其他取消操作
|
||||
- 完整展示执行过程的日志和最终结果
|
||||
|
||||
> 💡 此示例适用于需要人工监督的场景，如金融交易、系统配置修改等高风险操作。
|
||||
|
||||
---
|
||||
|
||||
### 2. `02_graph_as_a_tool.py` （子图工具化封装）
|
||||
|
||||
演示如何将一个完整的 LangGraph 子工作流封装为单个工具（Graph-as-a-Tool 模式）。
|
||||
|
||||
- ✅ 掌握点：
|
||||
- 定义独立的子图状态结构（如 `RetryState`）
|
||||
- 构建包含重试逻辑的子图工作流
|
||||
- 使用 `@tool` 装饰器将子图封装为可调用工具
|
||||
- 在主工作流中像使用普通工具一样调用子图
|
||||
|
||||
- 效果：
|
||||
- 子图内部实现了失败重试逻辑，对调用者完全透明
|
||||
- 主工作流可以无缝集成复杂的子流程，保持代码整洁
|
||||
- 模拟实现了不稳定 API 调用的容错处理
|
||||
|
||||
> 💡 这种模式特别适合封装具有复杂内部逻辑的操作，如网络请求、数据库事务等可能需要重试或错误处理的场景。
|
||||
|
||||
---
|
||||
|
||||
### 3. `03_multi_agent_orchestration.py` （多智能体协作系统）
|
||||
|
||||
构建由总控节点协调多个专家智能体的协作系统，实现专业化分工。
|
||||
|
||||
- ✅ 掌握点：
|
||||
- 定义共享的多智能体状态结构（`AgentState`）
|
||||
- 创建不同功能的专家节点（RAG专家、网络研究员、代码编写者）
|
||||
- 实现 Supervisor 总控节点，负责任务分配和流程协调
|
||||
- 为每个专家配置专属工具和系统指令
|
||||
- 设计复杂的条件路由逻辑，支持多节点协作
|
||||
|
||||
- 效果：
|
||||
- 系统会根据用户问题自动选择最合适的专家处理
|
||||
- 专家可以调用专业工具获取信息
|
||||
- 总控节点监控对话进度，决定是否完成任务或切换专家
|
||||
- 完整展示多智能体交互过程和决策路径
|
||||
|
||||
> 💡 此架构适用于需要多领域专业知识协作解决的复杂问题，如技术咨询、客户服务等场景。
|
||||
|
||||
---
|
||||
|
||||
### 4. `04_supervisor_agent_ops_system.py` （企业级运维智能体系统）
|
||||
|
||||
整合前述所有高级特性，构建一个完整的企业级 IT 运维智能体系统。
|
||||
|
||||
- ✅ 掌握点：
|
||||
- **Graph-as-a-Tool**：封装 SSH 日志分析子图，处理连接重试逻辑
|
||||
- **Human-in-the-Loop**：对重启服务等高危操作实施人工审批
|
||||
- **Multi-Agent 协作**：实现日志专家和运维专家的专业分工
|
||||
- **Supervisor 调度**：根据问题类型和解决进度智能分配任务
|
||||
- **状态持久化**：使用 MemorySaver 维护跨轮次的会话状态
|
||||
|
||||
- 效果：
|
||||
- 完整模拟 IT 故障处理流程：报警 → 日志分析 → 故障定位 → 人工审批 → 执行修复
|
||||
- 自动化程度高，但保留人工干预的安全机制
|
||||
- 各组件松耦合，易于扩展和维护
|
||||
- 支持 LangSmith 全流程追踪和调试
|
||||
|
||||
> 💡 这是一个接近生产环境的综合示例，展示了 LangGraph 构建复杂智能体系统的强大能力。
|
||||
|
||||
---
|
||||
|
||||
### 🔔 全局注意事项
|
||||
|
||||
- **学习路径建议**：
|
||||
`01`（人工干预） → `02`（子图封装） → `03`（多智能体基础） → `04`（企业级综合应用）
|
||||
- 所有 `.py` 文件均依赖根目录 `.env` 中的 `OPENAI_API_KEY` 和 `LANGCHAIN_API_KEY`
|
||||
- LangSmith 追踪功能已默认开启，可在 [LangSmith UI](https://smith.langchain.com) 实时查看执行流程
|
||||
- `04_supervisor_agent_ops_system.py` 是一个整合性示例，建议在理解前三个文件的基础上学习
|
||||
|
||||
---
|
||||
|
||||
### 💡 **建议**
|
||||
- 尝试扩展 `04_supervisor_agent_ops_system.py`，添加更多专业角色（如安全专家、性能分析师）
|
||||
- 实验不同的 Supervisor 决策逻辑，优化任务分配策略
|
||||
- 将真实的运维工具或 API 集成到系统中，构建实用的运维助手
|
||||
- 探索在生产环境中部署此类系统的最佳实践，如性能优化、错误处理和监控
|
||||
- 考虑添加更复杂的工作流模式，如并行执行、超时处理等高级特性
|
||||
@@ -28,10 +28,10 @@
|
||||
## 🧭 模块总览（16 模块）
|
||||
|
||||
> 💡 每个模块对应独立目录（含代码、说明与示例），可独立运行与学习。
|
||||
> 🔥 已更新至 08 LangGraph 基础篇，持续更新中！
|
||||
> 🔥 已更新至 09 LangGraph 进阶篇，持续更新中！
|
||||
|
||||
| 章节 | 模块 | 博客 | 核心关键词 | 难度 |
|
||||
|:-----------:|:--------------------------------------------------------------------------------------------------------|:-------------------------------------------------------------------:|:----------------------------------------------------|:-----:|
|
||||
|:-----------:|:--------------------------------------------------------------------------------------------------------|:-------------------------------------------------------------------:|:-------------------------------------------------------------|:-----:|
|
||||
| 🏗️ **基础篇** | [01 Agent 入门 & 环境搭建](https://github.com/Annyfee/agent-craft/tree/main/01_agent_introduction) | [🏠](https://blog.csdn.net/2401_87328929/article/details/153729555) | OpenAI API | ⭐ |
|
||||
| | [02 LLM 基础调用](https://github.com/Annyfee/agent-craft/tree/main/02_llm_fundamentals) | [🏠](https://blog.csdn.net/2401_87328929/article/details/153735431) | LLM API 调用 · prompt · 上下文记忆 | ⭐ |
|
||||
| | [03 Function Calling 与工具调用](https://github.com/Annyfee/agent-craft/tree/main/03_function_calling_tools) | [🏠](https://blog.csdn.net/2401_87328929/article/details/153866573) | Function Call · 工具函数封装 | ⭐⭐ |
|
||||
@@ -40,8 +40,8 @@
|
||||
| | [06 RAG 基础篇](https://github.com/Annyfee/agent-craft/tree/main/06_rag_basics) | [🏠](https://blog.csdn.net/2401_87328929/article/details/154230067) | RAG概念 · Split · Embedding · FAISS · RAG 链 | ⭐⭐ |
|
||||
| | [07 RAG 进阶篇](https://github.com/Annyfee/agent-craft/tree/main/07_rag_advanced) | [🏠](https://blog.csdn.net/2401_87328929/article/details/154408744) | Chroma · Reranker · RAG工具化 · 集成 | ⭐⭐⭐ |
|
||||
| | [08 LangGraph 基础篇](https://github.com/Annyfee/agent-craft/tree/main/08_langgraph_basics) | [🏠](https://blog.csdn.net/2401_87328929/article/details/154576812) | Langgraph三要素 · LangSmith · ReAct · 提示词安全注入 | ⭐⭐⭐ |
|
||||
| | 09 LangGraph 进阶篇 | 🚧撰写中 | `Multi-Agent` · `Human-in-the-Loop` · “总控-员工”架构 | ⭐⭐⭐⭐ |
|
||||
| 🧠 **智能篇** | 10 MCP 基础 (Client) | 🚧 | `MCP` 协议 · `CherryStudio` · 消费公共 MCP | ⭐⭐⭐ |
|
||||
| | [09 LangGraph 进阶篇]() | [🏠]() | Human-in-the-Loop · Graph-as-a-Tool · Multi-Agent 多智能体编排 | ⭐⭐⭐⭐ |
|
||||
| 🧠 **智能篇** | 10 MCP 基础 (Client) | 🚧撰写中 | `MCP` 协议 · `CherryStudio` · 消费公共 MCP | ⭐⭐⭐ |
|
||||
| | 11 MCP 进阶 (Server) | 🚧 | `Streamable HTTP` · `src layout` · 构建私有 MCP | ⭐⭐⭐⭐ |
|
||||
| | 12 Swarm & Agents SDK | 🚧 | `Swarm` · `Handoff` · “航空公司客服”项目 | ⭐⭐⭐⭐ |
|
||||
| 🏁 **实战篇** | 13 Streamlit 快速入门 | 🚧 | `st.chat_input` · `st.session_state` · 封装 RAG Agent | ⭐⭐⭐ |
|
||||
@@ -53,7 +53,7 @@
|
||||
|
||||
## 🧩 当前学习建议（适用于已完成模块）
|
||||
|
||||
目前已开放模块：**01 ~ 08**。
|
||||
目前已开放模块：**01 ~ 09**。
|
||||
建议按顺序学习，每一章都有完整代码示例与注释。
|
||||
|
||||
### ✅ 模块 01 — Agent 入门 & 环境搭建
|
||||
@@ -96,7 +96,12 @@
|
||||
- **目标**：懂得Langgraph的三要素，会使用LangSmith调试当前agent，并能自己写一个简单的Langgraph流程。
|
||||
- **内容**：Langgraph核心三要素｜ReAct循环白盒化实现 | LangSmith调试追踪 | 持久化记忆 | 提示词安全注入
|
||||
|
||||
> 📌 后续模块将陆续开放（MCP、多智能体、Streamlit等）
|
||||
### ✅ 模块 09 — LangGraph 进阶篇
|
||||
|
||||
- **目标**：学会人工审批机制的添加，子图工具化的封装与多智能体流程编排。最终可以融合这三个知识点做一个复杂Agent出来。
|
||||
- **内容**：Human-in-the-Loop（人工干预）｜ Graph-as-a-Tool（图即工具）| Multi-Agent 多智能体编排
|
||||
|
||||
> 📌 后续模块将陆续开放（MCP、Streamlit、Swarm等）
|
||||
|
||||
---
|
||||
|
||||
|
||||
Reference in New Issue
Block a user