大模型Function Calling实战指南：从原理到代码，让AI更强大

本文详解大模型函数调用(Function Calling)技术，包括核心概念、与ReACT的区别、工具定义格式及应用场景。通过Python代码示例展示如何让大模型执行计算任务，获取更准确结果。Function Calling使大模型能与外部服务交互，适用于API调用、数据库查询等场景，是提升AI应用能力的重要技术。

Function/Tool Calling

@openai: 函数调用为 OpenAI 模型提供了一种强大而灵活的方式，可以与您的代码或外部服务进行交互。通过函数调用，您可以向 Assistant 描述函数，并让其智能地返回需要调用的函数及其参数。@openai
@langchain: 我们将“工具调用”与“函数调用”互换使用。虽然“函数调用”有时指的是单个函数的调用，但我们将所有模型视为可以在每条消息中返回多个工具或函数调用。

工具调用与 ReACT 的区别

项目	ReAct 提示词工程	Function Calling
核心思想	用提示词引导模型「推理 + 行动 + 观察反馈」循环	明确声明函数签名，模型调用结构化函数
技术机制	Prompt 模板控制流程，如：Think -> Act -> Obs	通过函数描述注册函数接口，模型自动生成调用参数
环境交互灵活性	高：适用于复杂多步任务、agent 类流程	中：适合明确定义的单步工具调用
开发复杂度	中：需要精心设计提示词模板与中间状态管理	低~中：注册函数较直接，但需要参数与返回值处理
典型应用场景	多步推理、多工具交替使用、信息抽取等复杂场景	明确 API 调用、数据库查询、插件系统等
与 Agent 框架集成度	高：常用于 LangChain、AutoGPT 中的推理行为生成	高：OpenAI、LangChain、Claude 支持直接注册使用
优缺点总结	✅ 灵活可控	✅ 自动结构化调用
❌ 难以规模化自动解析	❌ 不擅长多步控制与反馈处理

function calling 核心概念

tools: 函数列表
function: 单个函数定义
function.name: 函数名字
function.description: 函数作用描述
parameters: 函数形参说明

"tools": [ { "type": "function", "function": { "name": "current_time", "description": "A tool for getting the current time.", "parameters": { "properties": {}, "required": [], "type": "object" } } }, { "type": "function", "function": { "name": "simple_code", "description": "A tool for running code and getting the result back. Only native packages are allowed, network/IO operations are disabled. and you must use print() or console.log() to output the result or result will be empty.", "parameters": { "properties": { "code": { "description": "code to be executed, only native packages are allowed, network/IO operations are disabled.", "type": "string" }, "language": { "description": "language of the code, only \"python3\" and \"javascript\" are supported", "type": "string" } }, "required": ["language", "code"], "type": "object" } } }]

工具调用的返回格式

tool_calls: 工具调用序列，单个或者多个
function 函数调用
function.name 函数名字
function.arguments 函数实参

"message": { "role": "assistant", "content": "", "tool_calls": [ { "function": { "name": "simple_code", "arguments": { "code": "print(0.9111 ** 3)", "language": "python3" } } } ] }

常用工具

代码解析器 python node
命令解析器 bash
文件管理 file directory
浏览器控制 browser use
外部接口 openapi swagger
大模型上下文协议 MCP playwright-mcp

python 工具调用示例

没有工具调用的对话模型

使用工具可以获得更加准确的结果

带有 function call 的请求

{ "model": "qwen3","stream": false,"options": { "temperature": 0 },"messages": [ { "role": "system", "content": "/no_think\n" }, { "role": "user", "content": "使用python计算 0.9111**3=" } ],"tools": [ { "type": "function", "function": { "name": "current_time", "description": "A tool for getting the current time.", "parameters": { "properties": {}, "required": [], "type": "object" } } }, { "type": "function", "function": { "name": "simple_code", "description": "A tool for running code and getting the result back. Only native packages are allowed, network/IO operations are disabled. and you must use print() or console.log() to output the result or result will be empty.", "parameters": { "properties": { "code": { "description": "code to be executed, only native packages are allowed, network/IO operations are disabled.", "type": "string" }, "language": { "description": "language of the code, only \"python3\" and \"javascript\" are supported", "type": "string" } }, "required": ["language", "code"], "type": "object" } } } ]}

大模型的响应

{ "model": "qwen3","created_at": "2025-05-05T16:57:36.016688Z","message": { "role": "assistant", "content": "", "tool_calls": [ { "function": { "name": "simple_code", "arguments": { "code": "print(0.9111 ** 3)", "language": "python3" } } } ] },"done_reason": "stop","done": true,"total_duration": 2037533417,"load_duration": 34440000,"prompt_eval_count": 275,"prompt_eval_duration": 776227166,"eval_count": 42,"eval_duration": 1225024292}

最终请求

{ "model": "qwen3","stream": false,"options": { "temperature": 0 },"messages": [ { "role": "system", "content": "/no_think\n" }, { "role": "user", "content": "使用python计算 0.9111**3=" }, { "role": "assistant", "content": "" }, { "role": "tool", "content": "0.756307034631\n" } ],"tools": [ { "type": "function", "function": { "name": "current_time", "description": "A tool for getting the current time.", "parameters": { "properties": {}, "required": [], "type": "object" } } }, { "type": "function", "function": { "name": "simple_code", "description": "A tool for running code and getting the result back. Only native packages are allowed, network/IO operations are disabled. and you must use print() or console.log() to output the result or result will be empty.", "parameters": { "properties": { "code": { "description": "code to be executed, only native packages are allowed, network/IO operations are disabled.", "type": "string" }, "language": { "description": "language of the code, only \"python3\" and \"javascript\" are supported", "type": "string" } }, "required": ["language", "code"], "type": "object" } } } ]}

最终响应

{ "model": "qwen3","created_at": "2025-05-05T16:57:37.121945Z","message": { "role": "assistant", "content": "<think>\n\n</think>\n\n0.9111 raised to the power of 3 is approximately **0.7563**." },"done_reason": "stop","done": true,"total_duration": 1030672292,"load_duration": 11999667,"prompt_eval_count": 303,"prompt_eval_duration": 122764292,"eval_count": 29,"eval_duration": 892274041}

大模型Function Calling实战指南：从原理到代码，让AI更强大

Function/Tool Calling

工具调用与 ReACT 的区别

function calling 核心概念

工具调用的返回格式

常用工具

python 工具调用示例

没有工具调用的对话模型

使用工具可以获得更加准确的结果

带有 function call 的请求

大模型的响应

最终请求

最终响应

相关文章