秒懂Function Call：DeepSeek大模型的隐藏超能力，让AI从"会聊天"变"会做事" 原创

# 定义一个天气查询函数
def get_weather(location, date):
    """查询指定地点和日期的天气


    参数:
        location (str): 城市名称
        date (str): 日期，格式为YYYY-MM-DD


    返回:
        dict: 包含天气信息的字典
    """
    # 实际应用中，这里会调用天气API
    import requests


    api_url = f"https://weather-api.example.com/forecast?location={location}&date={date}"
    response = requests.get(api_url)


    if response.status_code == 200:
        return response.json()
    else:
        return {"error": "无法获取天气信息"}


# DeepSeek如何使用Function Call
def deepseek_with_function_call(user_query):
    # 1. 模型理解用户查询天气的意图
    if "天气" in user_query:
        # 2. 提取地点和时间信息
        location = extract_location(user_query)  # 假设这个函数能提取位置
        date = extract_date(user_query)  # 假设这个函数能提取日期


        # 3. 调用天气函数
        weather_data = get_weather(location, date)


        # 4. 格式化回复
        response = f"{location}{date}天气为{weather_data['condition']}，"
        response += f"温度在{weather_data['min_temp']}°C至{weather_data['max_temp']}°C之间，"
        response += f"降水概率{weather_data['precipitation_chance']}%。"


        return response


    # 其他类型的查询...
    return "我无法理解您的问题。"

# 商品搜索函数
def search_products(product_type, color=None, max_price=None, min_price=None):
    """搜索符合条件的商品


    参数:
        product_type (str): 商品类型
        color (str, optional): 颜色
        max_price (float, optional): 最高价格
        min_price (float, optional): 最低价格


    返回:
        list: 符合条件的商品列表
    """
    # 实际中会调用电商平台API
    # 这里用模拟数据演示
    all_products = [
        {"id": 1, "name": "夏季红色连衣裙", "color": "红色", "price": 259.00},
        {"id": 2, "name": "优雅红色长裙", "color": "红色", "price": 329.00},
        {"id": 3, "name": "简约红色连衣裙", "color": "红色", "price": 199.00}
    ]


    results = []
    for product in all_products:
        if product_type in product["name"] and \
           (color is None or product["color"] == color) and \
           (max_price is None or product["price"] <= max_price) and \
           (min_price is None or product["price"] >= min_price):
            results.append(product)


    return results


# DeepSeek处理购物查询
def handle_shopping_query(query):
    # 提取查询信息
    if "连衣裙" in query and "红色" in query:
        # 解析价格条件
        max_price = 300  # 从查询中提取"低于300元"


        # 调用商品搜索函数
        products = search_products("连衣裙", color="红色", max_price=max_price)


        # 格式化回复
        if products:
            response = f"我找到了{len(products)}款符合条件的红色连衣裙：\n"
            for p in products:
                response += f"- {p['name']}，售价{p['price']}元\n"
        else:
            response = "抱歉，没有找到符合条件的商品。"


        return response

# 注册可用函数
available_functions = {
    "get_weather": {
        "description": "获取指定地点和日期的天气信息",
        "parameters": {
            "type": "object",
            "properties": {
                "location": {
                    "type": "string",
                    "description": "城市名称，如'北京'、'上海'"
                },
                "date": {
                    "type": "string",
                    "description": "日期，格式为YYYY-MM-DD"
                }
            },
            "required": ["location"]
        }
    },
    "search_products": {
        "description": "搜索符合条件的商品",
        "parameters": {
            "type": "object",
            "properties": {
                "product_type": {
                    "type": "string",
                    "description": "商品类型，如'连衣裙'、'手机'"
                },
                "color": {
                    "type": "string",
                    "description": "商品颜色"
                },
                "max_price": {
                    "type": "number",
                    "description": "最高价格"
                }
            },
            "required": ["product_type"]
        }
    }
}

def process_user_query(user_query):
    # 模拟DeepSeek分析用户意图的过程


    # 假设这是模型的思考过程
    """
    用户问"上海明天天气怎么样"，明显是查询天气信息。
    需要的信息：城市=上海，日期=明天(2025-02-27)
    应该调用get_weather函数
    """


    # 模型决定调用get_weather函数
    function_call = {
        "name": "get_weather",
        "arguments": {
            "location": "上海",
            "date": "2025-02-27"
        }
    }


    return function_call

def deepseek_complete_response(user_query):
# 1. 确定需要调用的函数
    function_call = process_user_query(user_query)


# 2. 执行函数调用
if function_call["name"] == "get_weather":
        args = function_call["arguments"]
        weather_data = get_weather(args["location"], args["date"])


# 3. 将函数结果融入自然语言回答
        response = f"{args['location']}在{args['date']}的天气预报：\n"
        response += f"天气状况: {weather_data['condition']}\n"
        response += f"温度范围: {weather_data['min_temp']}°C - {weather_data['max_temp']}°C\n"
        response += f"降水概率: {weather_data['precipitation_chance']}%\n"


if weather_data['precipitation_chance'] > 50:
            response += "建议您出门携带雨伞！"


return response


# 其他函数调用类似处理...

# 传统API调用 - 开发者必须硬编码逻辑
def traditional_weather_app(user_input):
    if "天气" in user_input and extract_location(user_input):
        location = extract_location(user_input)
        date = extract_date(user_input) or "today"
        try:
            weather_data = weather_api.get_forecast(location, date)
            return format_weather_response(weather_data)
        except ApiError as e:
            return f"获取天气失败: {e}"
    else:
        return "我不理解您的问题"


# Function Call - 模型自主决策
def function_call_assistant(user_input):
    # 模型分析输入，自主决定是否需要天气信息
    function_decision = model.analyze(user_input)


    if function_decision.needs_function_call:
        # 模型自动生成结构化参数
        function_name = function_decision.function_name  # 可能是"get_weather"或其他
        function_args = function_decision.arguments  # 模型提取的参数


        try:
            result = call_function(function_name, function_args)
        except Exception as e:
            # 模型能理解错误并尝试修正参数后重试
            corrected_args = model.correct_arguments(function_args, error=str(e))
            result = call_function(function_name, corrected_args)


        # 模型融合函数结果生成回答
        return model.generate_response(user_input, function_result=result)
    else:
        # 模型判断不需要调用外部函数
        return model.generate_response(user_input)