Tool&Function Calling

发表于 更新于 分类于
本文字数： 9.7k 阅读时长 ≈ 18 分钟

概述

Function Calling 是 OpenAI 最早提出的一项重要功能，旨在让大型语言模型（LLM）能够调用外部工具或服务，后续其他LLM厂商都逐渐支持这项功能。其实现方式是：通过定义一组标准化的 JSON 格式接口，使模型能够理解并生成对特定函数的调用请求。这些外部工具可以包括网页搜索、数据库查询、文件解析等通用服务，也可以是用户自定义的业务逻辑函数。

借助 Function Calling，LLM 可以将原本无法处理的任务“委托”给相应的工具执行，从而显著提升其实际应用能力和推理深度。这种机制不仅增强了模型的实用性，也为构建基于 LLM 的复杂系统提供了良好的扩展基础。

在后续的发展中，这一功能逐渐被更广泛地称为 Tool Calling，两者语义等价。

Function calling

对于 Function calling 的使用流程，通常可以划分为以下五个步骤：

  1. 定义可用工具：在向 LLM 发送请求前，需在请求参数中明确声明可供调用的工具及其功能描述，使模型了解它可以借助哪些外部能力来完成任务；
  2. LLM 生成工具调用指令：模型在理解用户问题后，结合已知的可用工具，决定是否需要调用某个或多个工具，并输出对应的工具名称及参数；
  3. 执行工具调用：根据模型返回的调用信息，系统实际调用相应的外部工具（如 API、数据库查询等），并传入指定参数以获取真实数据；
  4. 注入工具结果至上下文：将工具返回的结果整合进原始对话上下文中，作为新的输入再次传递给 LLM，供其进行进一步的理解和推理；
  5. LLM 生成最终自然语言响应：模型基于用户的原始提问和工具返回的数据，生成最终的自然语言回复，完成整个交互过程。

注意：一次完整的 Function calling 流程包含两次 LLM 推理调用：

  • 第一次调用：输入为用户提问 + 可用工具描述，输出为模型希望调用的工具及参数。
  • 第二次调用：输入为用户提问 + 工具返回的实际结果，输出为最终面向用户的自然语言回答。

这一流程构成了 Function calling 的完整交互闭环，使得语言模型能够在自身能力之外，借助外部工具动态接入实时、外部的数据与服务，从而实现更强大和实用的功能。

第一步，定义函数get_weather，输入精度和维度就可以获取该地区的气温，并且定义好该函数的调用格式

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
import requests

def get_weather(latitude, longitude):
    response = requests.get(
        f"https://api.open-meteo.com/v1/forecast?latitude={latitude}&longitude={longitude}&current=temperature_2m,wind_speed_10m&hourly=temperature_2m,relative_humidity_2m,wind_speed_10m")
    data = response.json()
    return data['current']['temperature_2m']

tools = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get current temperature for provided coordinates in celsius.",
            "parameters": {
                "type": "object",
                "properties": {
                    "latitude": {
                        "type": "number",
                        "description": "The latitude",
                    },
                    "longitude": {
                        "type": "number",
                        "description": "The longitude",
                    },
                },
                "required": ["latitude", "longitude"],
            },
        },
    }
]

第二步：调用大模型，输入查询内容和可供调用工具

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
import litellm

messages = [
    {
        "role": "user",
        "content": "What's the weather like in Paris today?"
    }
]

response = litellm.completion(
    model="openai/gpt-4.1",
    messages=messages,
    tools=tools,
    tool_choice="auto",  # auto is default, but we'll be explicit
)
print("\nFirst LLM Response:\n")
print(response.model_dump_json(indent=2))
# OUTPUT
First LLM Response:

{
  "id": "chatcmpl-BksIxuE09LW7FhkuTf7vIH84r6EV5",
  "created": 1750511535,
  "model": "gpt-4.1-2025-04-14",
  "object": "chat.completion",
  "system_fingerprint": "fp_07e970ab25",
  "choices": [
    {
      "finish_reason": "tool_calls",
      "index": 0,
      "message": {
        "content": null,
        "role": "assistant",
        "tool_calls": [
          {
            "index": 0,
            "function": {
              "arguments": "{\"latitude\":48.8566,\"longitude\":2.3522}",
              "name": "get_weather"
            },
            "id": "call_NiNlG5wjgIqikZmS608qWYAo",
            "type": "function"
          }
        ],
        "function_call": null
      }
    }
  ],
  "usage": {
    "completion_tokens": 25,
    "prompt_tokens": 66,
    "total_tokens": 91,
    "completion_tokens_details": {
      "accepted_prediction_tokens": null,
      "audio_tokens": null,
      "reasoning_tokens": null,
      "rejected_prediction_tokens": null
    },
    "prompt_tokens_details": {
      "audio_tokens": null,
      "cached_tokens": null
    }
  },
  "service_tier": null
}

第三步：解析大模型输出的结果

1
2
3
4
5
6
7
response_message = response.choices[0].message
print(response_message)
tool_calls = response.choices[0].message.tool_calls
print(tool_calls)
# OUTPUT
Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(index=0, function=Function(arguments='{"latitude":48.8566,"longitude":2.3522}', name='get_weather'), id='call_8h7sJ2Yxn6E0cMAgs7yaxv7W', type='function')], function_call=None, provider_specific_fields={'refusal': None})
[ChatCompletionMessageToolCall(index=0, function=Function(arguments='{"latitude":48.8566,"longitude":2.3522}', name='get_weather'), id='call_8h7sJ2Yxn6E0cMAgs7yaxv7W', type='function')]

根据解析结果，调用相应的外部工具

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
import json

# Check if the model wants to call a function
if tool_calls:
    # Execute the functions and prepare responses
    available_functions = {
        "get_weather": get_weather,
    }

    messages.append(response_message)  # Extend conversation with assistant's reply

    for tool_call in tool_calls:
        print(f"\nExecuting tool call\n{tool_call}")
        function_name = tool_call.function.name
        function_to_call = available_functions[function_name]
        function_args = json.loads(tool_call.function.arguments)
        # calling the get_current_weather() function
        function_response = function_to_call(
            latitude=function_args.get("latitude"),
            longitude=function_args.get("longitude"),
        )
        print(f"Result from tool call\n{function_response}\n")

        # Extend conversation with function response
        messages.append(
            {
                "tool_call_id": tool_call.id,
                "role": "tool",
                "name": function_name,
                "content": str(function_response),
            }
        )
# OUTPUT
Executing tool call
ChatCompletionMessageToolCall(index=0, function=Function(arguments='{"latitude":48.8566,"longitude":2.3522}', name='get_weather'), id='call_8h7sJ2Yxn6E0cMAgs7yaxv7W', type='function')
Result from tool call
33.6

第四步/第五步：输入提问结合外部工具获取的结果，进行第二次大模型计算，生成最终结果

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
second_response = litellm.completion(
    model="openai/gpt-4.1",
    messages=messages,
)
print("Second Response\n", second_response.model_dump_json(indent=2))
message = second_response.choices[0].message.content
print(f"Message: {message}")
# OUTPUT
Second Response
 {
  "id": "chatcmpl-Bkskgv18D7JiQWLDeOPtRHJFQbplo",
  "created": 1750513254,
  "model": "gpt-4.1-2025-04-14",
  "object": "chat.completion",
  "system_fingerprint": "fp_07e970ab25",
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "message": {
        "content": "Today in Paris, the temperature is currently around 33.6°C. It's quite a hot day, so if you're heading outside, remember to stay hydrated and wear sun protection! If you need more details on weather conditions, let me know.",
        "role": "assistant",
        "tool_calls": null,
        "function_call": null
      }
    }
  ],
  "usage": {
    "completion_tokens": 50,
    "prompt_tokens": 50,
    "total_tokens": 100,
    "completion_tokens_details": {
      "accepted_prediction_tokens": null,
      "audio_tokens": null,
      "reasoning_tokens": null,
      "rejected_prediction_tokens": null
    },
    "prompt_tokens_details": {
      "audio_tokens": null,
      "cached_tokens": null
    }
  },
  "service_tier": null
}
Message: Today in Paris, the temperature is currently around 33.6°C. It's quite a hot day, so if you're heading outside, remember to stay hydrated and wear sun protection! If you need more details on weather conditions, let me know.

Parallel function calling

在某些高级应用场景中，大型语言模型（LLM）可以在一次推理过程中并行地调用多个工具函数。传统上，模型在处理需要外部数据的任务时，会按顺序依次调用函数，这可能导致较高的延迟。而LLM接口（如 OpenAI 的 tool_call 机制）支持在单次响应中返回多个函数调用指令，从而实现并行执行，显著缩短整体响应时间。

考虑以下用户输入：

“今天旧金山、东京和巴黎的天气分别怎么样？”

这是一个典型的多目标查询，要求模型获取三个不同城市的实时天气信息。为了高效完成该请求，LLM 可以在单次回复中并行调用三次 get_weather 函数，每个调用对应一个城市的位置参数。

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
# INPUT MESSAGES
messages = [
    {
        "role": "user",
        # "content": "What's the weather like in Paris today?"
        "content": "What's the weather like in San Francisco, Tokyo, and Paris today?"
    }
]
# Tool Calls
tool_calls = response.choices[0].message.tool_calls
pprint.pprint(tool_calls)
# OUTPUT
[ChatCompletionMessageToolCall(index=0, function=Function(arguments='{"latitude": 37.7749, "longitude": -122.4194}', name='get_weather'), id='call_BBuG4wEfDvmZ61eFdihdinDR', type='function'),
 ChatCompletionMessageToolCall(index=0, function=Function(arguments='{"latitude": 35.6895, "longitude": 139.6917}', name='get_weather'), id='call_O606wPtT2qMlqUtIBNVIrrwD', type='function'),
 ChatCompletionMessageToolCall(index=0, function=Function(arguments='{"latitude": 48.8566, "longitude": 2.3522}', name='get_weather'), id='call_nEqcSa7pZt9Jklni57JnReCi', type='function')]
# Second LLM
second_response = litellm.completion(
    model="openai/gpt-4.1",
    messages=messages,
    base_url=base_url,
)
message = second_response.choices[0].message.content
print(f"Message: {message}")
# OUTPUT
Message: Today's weather:

- San Francisco: Around 14°C, cool and typically foggy or breezy for this time of year.
- Tokyo: Around 30°C, warm and possibly humid.
- Paris: Around 23°C, mild and likely pleasant.

Let me know if you need more detailed weather info or forecasts!

相关阅读