1
用户："北京今天天气怎么样？"
2

3
LLM的思考：
4
1. 用户想查天气
5
2. 我有个天气查询工具
6
3. 需要的参数是城市名
7
4. 用户说的是"北京"
8
5. 决定：调用天气工具，参数city="北京"
9

10
返回：
11
{
12
  "function": "get_weather",
13
  "arguments": {"city": "北京"}
14
}
15

16
你的代码执行真正的天气API调用
17
返回结果给LLM
18
LLM生成自然语言回答

1
tools = [
2
    {
3
        "type": "function",
4
        "function": {
5
            "name": "get_weather",
6
            "description": "获取指定城市的当前天气",
7
            "parameters": {
8
                "type": "object",
9
                "properties": {
10
                    "city": {
11
                        "type": "string",
12
                        "description": "城市名称，如：北京、上海"
13
                    }
14
                },
15
                "required": ["city"]
16
            }
17
        }
18
    }
19
]

1
tools = [
2
    {
3
        "type": "function",
4
        "function": {
5
            "name": "get_weather",
6
            "description": "获取指定城市的天气信息",
7
            "parameters": {
8
                "type": "object",
9
                "properties": {
10
                    "city": {"type": "string", "description": "城市名"}
11
                },
12
                "required": ["city"]
13
            }
14
        }
15
    }
16
]

1
response = client.chat.completions.create(
2
    model="gpt-4o",
3
    messages=[
4
        {"role": "user", "content": "北京今天天气怎么样？"}
5
    ],
6
    tools=tools
7
)

1
# LLM返回工具调用决策
2
tool_call = response.choices[0].message.tool_calls[0]
3

4
if tool_call.function.name == "get_weather":
5
    args = json.loads(tool_call.function.arguments)
6
    # args = {"city": "北京"}
7

8
    # 执行真正的天气查询
9
    weather = get_weather(args["city"])
10
    # weather = {"temp": 25, "condition": "晴"}

1
# 把天气结果返回给LLM
2
messages.append({
3
    "role": "tool",
4
    "content": json.dumps(weather)
5
})
6

7
# LLM生成最终回答
8
final_response = client.chat.completions.create(
9
    model="gpt-4o",
10
    messages=messages
11
)
12
# "北京今天天气晴朗，气温25度。"

1
tools = [
2
    {
3
        "type": "function",
4
        "function": {
5
            "name": "search_web",
6
            "description": "搜索互联网获取信息"
7
        }
8
    },
9
    {
10
        "type": "function",
11
        "function": {
12
            "name": "query_database",
13
            "description": "查询公司内部数据库"
14
        }
15
    },
16
    {
17
        "type": "function",
18
        "function": {
19
            "name": "send_email",
20
            "description": "发送邮件"
21
        }
22
    }
23
]

1
用户："帮我查一下最近销售数据，然后发邮件给销售总监"
2

3
AI的决策：
4
Step 1: 调用 query_database
5
Step 2: 调用 send_email
6

7
AI会自动规划工具调用顺序

1
 "查询天气"
2
 "获取指定城市的当前天气信息，包括温度、湿度、天气状况"

1
"parameters": {
2
    "properties": {
3
        "city": {"type": "string"},
4
        "unit": {
5
            "type": "string",
6
            "enum": ["celsius", "fahrenheit"]  # 限制可选值
7
        }
8
    }
9
}

1
try:
2
    result = execute_tool(tool_name, args)
3
except Exception as e:
4
    # 把错误返回给LLM，让它调整策略
5
    return f"工具执行失败：{str(e)}"

1
太粗：一个工具做太多事
2
太细：工具太多，AI难以选择
3

4
原则：一个工具做一件事