from openclaw import Agent, tool

@tool
def search_docs(query: str) -> str:
    """搜索知识库文档"""
    # 这里接入你的搜索引擎
    return search_engine.search(query)

@tool
def generate_answer(question: str, context: str) -> str:
    """根据上下文生成回答"""
    return llm.generate(f"问题：{question}\n上下文：{context}")

agent = Agent(
    model="gpt-4o",
    tools=[search_docs, generate_answer],
    system_prompt="你是一个知识库助手。先搜索文档，再根据结果回答问题。",
    max_retries=3,
)

result = agent.run("OpenClaw和LangGraph有什么区别？")
print(result)

from crewai import Agent, Task, Crew, Process

researcher = Agent(
    role="文档搜索专家",
    goal="找到和用户问题最相关的文档片段",
    backstory="你是一个经验丰富的文档检索工程师",
    tools=[search_docs],
    llm="gpt-4o",
)

writer = Agent(
    role="技术文档撰写者",
    goal="根据搜索结果写出准确、清晰的回答",
    backstory="你是一个技术写作专家，擅长用简洁的语言解释复杂概念",
    llm="gpt-4o",
)

search_task = Task(
    description="搜索关于'OpenClaw和LangGraph区别'的文档",
    expected_output="相关的文档片段列表",
    agent=researcher,
)

answer_task = Task(
    description="根据搜索结果，回答用户的问题",
    expected_output="一段200字左右的技术对比说明",
    agent=writer,
    context=[search_task],
)

crew = Crew(
    agents=[researcher, writer],
    tasks=[search_task, answer_task],
    process=Process.sequential,
    max_retries=3,
)

result = crew.kickoff()
print(result)

from langgraph.graph import StateGraph, END
from typing import TypedDict, List

class AgentState(TypedDict):
    question: str
    search_results: List[str]
    answer: str
    retry_count: int

def search_node(state: AgentState) -> AgentState:
    results = search_docs(state["question"])
    return {**state, "search_results": results}

def generate_node(state: AgentState) -> AgentState:
    context = "\n".join(state["search_results"])
    answer = generate_answer(state["question"], context)
    return {**state, "answer": answer, "retry_count": state["retry_count"] + 1}

def should_retry(state: AgentState) -> str:
    if is_answer_good(state["answer"]) or state["retry_count"] >= 3:
        return "end"
    return "search"

graph = StateGraph(AgentState)
graph.add_node("search", search_node)
graph.add_node("generate", generate_node)

graph.set_entry_point("search")
graph.add_edge("search", "generate")
graph.add_conditional_edges("generate", should_retry, {
    "search": "search",
    "end": END,
})

app = graph.compile()
result = app.invoke({
    "question": "OpenClaw和LangGraph有什么区别？",
    "search_results": [],
    "answer": "",
    "retry_count": 0,
})
print(result["answer"])

from autogen import AssistantAgent, UserProxyAgent

assistant = AssistantAgent(
    name="知识库助手",
    system_message="""你是一个知识库助手。当用户提问时：
    1. 先调用search_docs工具搜索相关文档
    2. 根据搜索结果生成回答
    3. 如果你认为回答不够完整，可以重新搜索""",
    llm_config={"model": "gpt-4o"},
    max_consecutive_auto_reply=3,
)

user_proxy = UserProxyAgent(
    name="用户",
    human_input_mode="NEVER",
    code_execution_config=False,
)

user_proxy.initiate_chat(
    assistant,
    message="OpenClaw和LangGraph有什么区别？",
)

你的场景是什么？
├── 单Agent，功能明确
│   ├── 需要高性能 → OpenClaw
│   └── 需要快速原型 → OpenClaw
├── 多Agent协作
│   ├── 角色分工明确 → CrewAI
│   ├── 需要复杂状态管理 → LangGraph
│   └── 需要自由对话式协作 → AutoGen
├── 复杂业务流程
│   ├── 有审批/回溯需求 → LangGraph
│   └── 有并行处理需求 → LangGraph
├── 企业级部署
│   ├── 基础设施在Azure上 → AutoGen
│   ├── 需要完整监控 → LangGraph + LangSmith
│   └── 需要细粒度权限控制 → OpenClaw
└── 预算有限
    ├── 开源方案 → OpenClaw 或 CrewAI
    └── 可以付费 → LangGraph + LangSmith