像你告诉你的朋友你需要在6点到机场,他们开始告诉你他们最喜欢的机场餐厅吗?这基本上是编程在功能呼叫之前的样子。 什么是函数呼叫? 函数调用是当人工智能理解它需要使用外部工具或服务来完成一个任务时,这就像拥有一个智能助理,他不仅了解你想要什么,而且知道何时拿出手机来检查天气,寻找方向,或叫出租车。 In technical terms, function calling lets AI models identify when to use specific functions, what parameters to pass, and how to format the data properly - all from natural language instructions. 《黑暗年代:功能召唤之前》 在函数调用之前,事情是......混乱的.如果你想要人工智能使用外部工具,你有两个同样令人沮丧的选择: 
 
 
 
 
 Train the AI to generate exact API calls (which often went wrong) 
 
 Write complex code to parse AI responses and extract parameters 你会说一些类似“我需要纽约的天气数据”,人工智能会用有关天气的文本回应,而不是你所需要的实际数据。 我记得建立了一个聊天机器人,该机器人应该预订餐厅预订。没有功能调用,它会很高兴地告诉用户,“我已经预订了你的桌子!”而没有真正连接到任何预订系统。 为什么函数叫事物 函数调用桥梁人类语言和计算机系统之间的差距。 
 
 
 
 
 
 It's more reliable - the AI structures data exactly as functions need it 
 
 It's more capable - an AI can access real-time data, control devices, or run calculations 
 
 It's more natural - users can make requests in plain language 想想函数调用就像给人工智能访问工具箱一样。以前,它只能谈论锤子,现在它可以拿起一个,实际上驱动一根钉子。 一个简单的方式来思考它 想象一下你在一家咖啡店,当你说,“我喜欢一杯中等的牛奶牛奶,”你基本上是“功能呼叫”酒吧。 这个功能可能看起来像: make_coffee(size="medium", drink_type="latte", milk="oat")
 酒吧人知道: 
 
 
 
 
 
 
 Identify the function needed (make a drink) 
 
 Extract the parameters (size, type, milk choice) 
 
 Execute the function (make the coffee) 
 
 Return the result (your drink) 没有函数调用,你会得到一个答案,如“一杯中等的牛奶和奶听起来很美味!”但没有真正的咖啡。 Hope you get the idea ! 函数呼叫背后的核心力学 在其核心,函数调用工作,因为几个技术创新一起工作: 
 
 
 
 
 
 
 
 : When you define functions for an AI model to use, you provide a clear schema that includes the function name, description, and parameter specifications with their types and requirements. This JSON schema helps the model understand not just what the function does, but exactly what information it needs to execute properly. Structured Function Definitions 
 
 : The AI model analyzes user input to determine when a function should be called. Modern models like GPT-4 can recognize when a user's request could be satisfied by calling an external function, even when the user doesn't explicitly ask for it. Intent Recognition 
 
 : The model identifies and extracts relevant parameters from natural language, converting them into the structured format required by the function. For example, when someone asks about Microsoft stock, the model knows to extract "MSFT" as the ticker parameter. Parameter Extraction 
 
 : After a function executes, the model receives the result and incorporates it into a natural language response, creating a seamless experience for the user. Response Formatting 
 
 : The model maintains context throughout the conversation, allowing it to use previous function calls and responses to inform new function calls. Context Awareness ️ Function Calling 函数调用代码 以下是您可以使用的模板来开始使用 OpenAI API 的函数调用: from openai import OpenAI
import json

# Initialize the client
client = OpenAI()

# Define your functions
functions = [
    {
        "type": "function",
        "function": {
            "name": "your_function_name",
            "description": "Description of what this function does",
            "parameters": {
                "type": "object",
                "properties": {
                    "param1": {
                        "type": "string",
                        "description": "Description of parameter 1"
                    },
                    "param2": {
                        "type": "integer",
                        "description": "Description of parameter 2"
                    }
                },
                "required": ["param1"]
            }
        }
    }
]

# Your actual implementation of the function
def your_function_name(param1, param2=None):
    # Your code here
    result = f"Did something with {param1}"
    if param2:
        result += f" and {param2}"
    return result

# Process messages and handle function calls
def process_with_function_calling(messages):
    # Get response from the model
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=messages,
        tools=functions,
        tool_choice="auto"
    )
    
    response_message = response.choices[0].message
    
    # Check if the model wants to call a function
    if response_message.tool_calls:
        # Append the model's response to messages
        messages.append(response_message)
        
        # Process each function call
        for tool_call in response_message.tool_calls:
            function_name = tool_call.function.name
            function_args = json.loads(tool_call.function.arguments)
            
            # Call the function
            if function_name == "your_function_name":
                function_response = your_function_name(
                    function_args.get("param1"),
                    function_args.get("param2")
                )
                
                # Append the function response to messages
                messages.append({
                    "role": "tool",
                    "tool_call_id": tool_call.id,
                    "name": function_name,
                    "content": function_response
                })
        
        # Get a new response from the model
        second_response = client.chat.completions.create(
            model="gpt-4o",
            messages=messages
        )
        
        return second_response.choices[0].message.content
    
    return response_message.content

# Example usage
messages = [{"role": "user", "content": "Your user query here"}]
result = process_with_function_calling(messages)
print(result)
 ☀️一个工作例子:天气检查器 让我们通过使用 Azure OpenAI 的简单天气检查功能来实现这一点: 

from openai import AzureOpenAI
import json
import requests


# Initialize the Azure OpenAI client
client = AzureOpenAI(
    api_key="your key",  
    api_version="2024-10-21",  # Make sure to use a version that supports function calling
    azure_endpoint="your endpoint"
)


# Define the weather function
weather_functions = [
    {
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get current weather for a location",
            "parameters": {
                "type": "object",
                "properties": {
                    "location": {
                        "type": "string",
                        "description": "City name, e.g., 'San Francisco'"
                    },
                    "unit": {
                        "type": "string",
                        "enum": ["celsius", "fahrenheit"],
                        "description": "Temperature unit"
                    }
                },
                "required": ["location"]
            }
        }
    }
]

# Implement the weather function
def get_weather(location, unit="celsius"):
    # In a real app, you'd call a weather API here
    # This is a mock implementation
    weather_data = {
        "San Francisco": {"temp": 18, "condition": "Foggy"},
        "New York": {"temp": 22, "condition": "Partly Cloudy"},
        "London": {"temp": 15, "condition": "Rainy"},
        "Tokyo": {"temp": 24, "condition": "Sunny"}
    }
    
    # Get data for the location or return a default
    data = weather_data.get(location, {"temp": 20, "condition": "Unknown"})
    
    # Convert temperature if needed
    temp = data["temp"]
    if unit == "fahrenheit":
        temp = (temp * 9/5) + 32
    
    return f"It's currently {data['condition']} and {temp}°{'F' if unit == 'fahrenheit' else 'C'} in {location}."

# Process messages with weather function
def check_weather(user_query):
    messages = [{"role": "user", "content": user_query}]
    
    # Get response from the model
    response = client.chat.completions.create(
        model="gpt-4o-mini",  # The name you gave your deployment
        messages=[{"role": "user", "content": user_query}],
        tools=weather_functions,
        tool_choice="auto"
    )
    
    response_message = response.choices[0].message
    
    # Check if the model wants to call the weather function
    if response_message.tool_calls:
        # Append the model's response to messages
        messages.append(response_message)
        
        # Process each function call
        for tool_call in response_message.tool_calls:
            function_name = tool_call.function.name
            function_args = json.loads(tool_call.function.arguments)
            
            # Call the weather function
            if function_name == "get_weather":
                function_response = get_weather(
                    function_args.get("location"),
                    function_args.get("unit", "celsius")
                )
                
                # Append the function response to messages
                messages.append({
                    "role": "tool",
                    "tool_call_id": tool_call.id,
                    "name": function_name,
                    "content": function_response
                })
        
        # Get a new response from the model
        second_response = client.chat.completions.create(
            model="gpt-4o-mini",
            messages=messages
        )
        
        return second_response.choices[0].message.content
    
    return response_message.content

# Try it out
result = check_weather("What's the weather like in Tokyo right now?")
print(result)
 当您运行此代码时,请问“现在东京的天气如何?”时,人工智能将: 
 
 
 
 
 
 
 Recognize it needs weather data 
 
 Call the   function with "Tokyo" as the location get_weather 
 
 Receive the weather data 
 
 Generate a natural response with the information 而不是猜测或编写天气信息,它通过功能提供真实的数据。 ️输出:天气函数呼吁行动 当我们运行我们的天气示例以查询“现在在东京的天气是什么样子?”时,呼叫函数的魔力产生了生命,该模型首先认识到这是一个天气请求,并构建了一个函数呼叫参数: 这个格式化的 JSON 被发送到我们的 该函数检索了我们定义的模糊天气数据:温度为24°C和“阳光”条件。 {"location": "Tokyo", "unit": "celsius"} get_weather 想要使用 OpenAI 代替 Azure OpenAI ? 两者之间的关键区别 和   实施的谎言 使用OpenAI,您使用标准客户端具有直接的API密钥,而Azure OpenAI则需要其特定的客户端以及其他参数,例如: 和 在执行函数调用时,OpenAI 将模型参考为其标准名称,而 Azure OpenAI 则要求您使用您在 Azure 门户中创建的部署名称。 OpenAI Azure OpenAI how you initialize the client azure_endpoint api_version 以下是代码比较: # Standard OpenAI implementation
from openai import OpenAI

# Initialize the client
client = OpenAI(
    api_key="your-openai-api-key"
)

# Make an API call
response = client.chat.completions.create(
    model="gpt-4o",  # Standard model name
    messages=[{"role": "user", "content": "What's the weather in Seattle?"}],
    tools=functions,
    tool_choice="auto"
)
 # Azure OpenAI implementation
from openai import AzureOpenAI

# Initialize the client - requires more parameters
client = AzureOpenAI(
    api_key="your-azure-openai-api-key",
    api_version="2023-07-01-preview",  # Azure-specific
    azure_endpoint="https://your-resource-name.openai.azure.com"  # Azure-specific
)

# Make an API call
response = client.chat.completions.create(
    model="your-deployed-model-name",  # The name you gave your deployment in Azure
    messages=[{"role": "user", "content": "What's the weather in Seattle?"}],
    tools=functions,
    tool_choice="auto"
)
 机遇的新世界 函数调用改变了一切,AI现在可以: 
 
 
 
 
 
 
 
 
 Book real appointments 
 
 Look up real-time data 
 
 Control smart home devices 
 
 Query databases 
 
 Process payments 
 
 And much more 这是一个可以谈论世界的人工智能和一个实际上可以与之互动的人之间的区别。 下次你问聊天机器人今天的新闻或打开你的灯,记住:这是一个功能的呼叫,使魔法发生在舞台后面。 六、总结 函数调用将人类语言和计算机系统之间的差距缩小,使人工智能能够了解何时使用正确格式化的数据调用外部工具和服务 - 无论您是直接使用OpenAI还是使用Azure OpenAI服务 - 我们已经探索了它如何将人工智能从被动响应器转变为能够根据自然语言请求采取具体行动的积极代理。 向前看 函数调用的真正力量不在于它今天能做些什么,而在于它明天能做些什么。随着越来越多的系统暴露了API终端,随着身份验证和安全标准的演变,我们正在走向一个自然语言成为数字系统的通用接口的世界。 最后的想法 真正的革命不仅在于人工智能现在可以采取行动,而在于思维和行动之间的障碍正在解散,函数调用不仅仅是一个技术特征,它是人工智能超越页面并进入世界的瞬间,从观察者转变为参与者。

This story contains new, firsthand information uncovered by the writer.

Books authored by me

Microsoft BUILD Session

PyCloud: Identifying defaulters using Machine Learning

Panel Discussion on Futuristic Technology and Possibilities

AI DEV DAYS: Integrating AI to Blockchain-Based Applications by Ritesh Modi

Understanding & Optimizing Dockerfiles by Ritesh Modi at Bangalore Container Conference 2017

Read My Stories

該音頻是用故事的原始語言製作的！

厌倦了破碎的聊天机器人?这个AI升级修复了一切

About Author

註釋

標籤

这篇文章刊登在

Related Stories

扬帆起航：利用数据湖开发生产级 RAG 应用程序

Claude Sonnet 3.5 系统提示泄漏：法医分析

HackerNoon Decoded: The Top 10 Countries Where HackerNoon Is the Most Active

架构师指南：构建 AI/ML 数据湖参考架构

扬帆起航：利用数据湖开发生产级 RAG 应用程序

Claude Sonnet 3.5 系统提示泄漏：法医分析

HackerNoon Decoded: The Top 10 Countries Where HackerNoon Is the Most Active

架构师指南：构建 AI/ML 数据湖参考架构

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps