Function Calling: LLM 函数调用

用户输入 → LLM 分析 → 返回函数调用指令 → 执行函数 → 返回结果 → LLM 生成自然语言响应

# 步骤1：意图识别
intent_prompt = "用户说：'明天北京会下雨吗'，意图是什么？返回：weather/email/calendar"
intent = llm.query(intent_prompt)  # 返回 "weather"

# 步骤2：参数提取
slot_prompt = "从'明天北京会下雨吗'中提取：城市、日期。返回JSON"
slots = json.loads(llm.query(slot_prompt))  # {"city": "北京", "date": "明天"}

# 步骤3：手动验证和调用
if intent == "weather":
    if "city" not in slots:
        return "请提供城市名称"
    result = get_weather(city=slots["city"], date=slots.get("date"))

# 一次性完成！
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "查询天气信息",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {
                        "type": "string",
                        "description": "城市名称"
                    },
                    "date": {
                        "type": "string",
                        "description": "日期",
                        "default": "今天"
                    }
                },
                "required": ["city"]
            }
        }
    }
]

response = openai.ChatCompletion.create(
    model="gpt-4",
    messages=[{"role": "user", "content": "明天北京会下雨吗"}],
    tools=tools,
    tool_choice="auto"
)

# LLM 自动返回：
# {
#   "tool_calls": [{
#     "function": {
#       "name": "get_weather",
#       "arguments": '{"city": "北京", "date": "明天"}'
#     }
#   }]
# }

# 直接执行
tool_call = response.choices[0].message.tool_calls[0]
result = get_weather(**json.loads(tool_call.function.arguments))

# 用户："查询明天北京的天气，并给张三发邮件通知"
# LLM 自动识别需要调用两个函数
response = {
    "tool_calls": [
        {
            "function": {
                "name": "get_weather",
                "arguments": '{"city": "北京", "date": "明天"}'
            }
        },
        {
            "function": {
                "name": "send_email",
                "arguments": '{"to": "zhangsan@example.com", "subject": "天气通知"}'
            }
        }
    ]
}

# 用户："那里明天会下雨吗？"（上下文：之前提到北京）
# LLM 自动从对话历史推断参数
response = {
    "tool_calls": [{
        "function": {
            "name": "get_weather",
            "arguments": '{"city": "北京", "date": "明天"}'  # 自动推断 city
        }
    }]
}

# 用户："提醒我3小时后开会"
# LLM 自动将"3小时后"转换为正确的时间格式
response = {
    "tool_calls": [{
        "function": {
            "name": "set_reminder",
            "arguments": '{"time": "2025-11-19T11:30:00", "message": "开会"}'
        }
    }]
}

import openai
import json

# 定义可用的工具
tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "获取指定城市的天气信息",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string", "description": "城市名称"},
                    "date": {"type": "string", "description": "日期，默认今天"},
                    "unit": {
                        "type": "string",
                        "enum": ["celsius", "fahrenheit"],
                        "description": "温度单位"
                    }
                },
                "required": ["city"]
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "get_weather_forecast",
            "description": "获取未来多天的天气预报",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string"},
                    "days": {"type": "integer", "description": "预报天数"}
                },
                "required": ["city", "days"]
            }
        }
    }
]

# 实际的工具函数
def get_weather(city, date="今天", unit="celsius"):
    # 实际调用天气 API
    return f"{city}{date}的天气：晴，25°C"

def get_weather_forecast(city, days):
    return f"{city}未来{days}天天气预报：..."

# Agent 主循环
def run_agent(user_message):
    messages = [{"role": "user", "content": user_message}]
    
    # 第一次调用：LLM 决定调用什么函数
    response = openai.ChatCompletion.create(
        model="gpt-4",
        messages=messages,
        tools=tools,
        tool_choice="auto"
    )
    
    response_message = response.choices[0].message
    
    # 如果 LLM 决定调用函数
    if response_message.tool_calls:
        # 执行函数调用
        for tool_call in response_message.tool_calls:
            function_name = tool_call.function.name
            function_args = json.loads(tool_call.function.arguments)
            
            # 调用实际函数
            if function_name == "get_weather":
                function_response = get_weather(**function_args)
            elif function_name == "get_weather_forecast":
                function_response = get_weather_forecast(**function_args)
            
            # 将函数结果添加到对话
            messages.append({
                "role": "tool",
                "tool_call_id": tool_call.id,
                "content": function_response
            })
        
        # 第二次调用：让 LLM 基于函数结果生成自然语言回复
        final_response = openai.ChatCompletion.create(
            model="gpt-4",
            messages=messages
        )
        
        return final_response.choices[0].message.content
    
    # 如果不需要调用函数，直接返回
    return response_message.content

# 测试
print(run_agent("明天北京会下雨吗？"))
# LLM 自动：1. 识别需要调用 get_weather
#          2. 提取参数 city="北京", date="明天"
#          3. 执行函数
#          4. 生成自然语言回复："根据查询，明天北京是晴天，不会下雨。"