LangGraph

from typing import Literal
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI
from langgraph.graph import END, StateGraph, START, MessagesState
from langgraph.prebuilt import ToolNode

# 1. 定义工具函数（模拟外部查询）
@tool
def search_weather(query: str):
    """查询指定城市的天气信息"""
    if "北京" in query.lower():
        return "阴天有雾，气温 25 度"
    return "天气晴朗，气温 35 度以上，注意防暑"

tools = [search_weather]
tool_node = ToolNode(tools)

# 2. 初始化适配 NoneLinear 的模型并绑定工具
llm = ChatOpenAI(
    model="DeepSeek-V3.2",
    api_key="YOUR_NONELINEAR_API_KEY",
    base_url="https://api.nonelinear.com/v1"
).bind_tools(tools)

# 3. 定义逻辑节点
def call_model(state: MessagesState):
    """节点：调用 LLM 获取响应"""
    response = llm.invoke(state["messages"])
    return {"messages": [response]}

def should_continue(state: MessagesState) -> Literal["tools", END]:
    """路由函数：判断是否需要继续调用工具"""
    last_message = state["messages"][-1]
    if last_message.tool_calls:
        return "tools"
    return END

# 4. 构建图结构
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)
workflow.add_edge("tools", "agent") # 工具执行完回到 Agent

# 5. 编译运行
app = workflow.compile()
inputs = {"messages": [("user", "帮我查一下北京的天气，然后给出穿衣建议。")]}

for chunk in app.stream(inputs, stream_mode="values"):
    chunk["messages"][-1].pretty_print()