GPT-4-0613调用函数时编造参数问题及解决咨询

问题说明 #

使用gpt-4-0613模型，配置了单个fill_form函数及自定义系统提示词。对话前两次触发函数时，API能正常向用户索要缺失的必填参数；但对话后期（如第5次请求）触发函数时，会编造对话上下文中不存在的虚构参数并返回函数调用结果。添加或移除function_call='auto'参数均无法解决该问题，需要找到阻止AI编造参数、始终向用户索要缺失参数的方法。

代码示例 #

completion = openai.ChatCompletion.create(
    model='gpt-4-0613',
    messages=prompts,
    functions=[
    {
        "name": "fill_form",
        "description": "Helps the user create an XYZ Report",
        "parameters": {
            "type": "object",
            "properties": {
                "name": {
                    "type": "string",
                    "description": "the full name of the person issuing this report"
                },
                "zip": {
                    "type": "string",
                    "description": "the 5 digit zip code of the address"
                },
                "address": {
                    "type": "string",
                    "description": "the street address, only the street and not the city, state or zip"
                },
                "year_end": {
                    "type": "string",
                    "description": "the full four digit year of the fiscal year"
                },
            },
            "required": ["name", "address", "year_end", "zip"]
        }
    }],
)

已尝试参数：

function_call='auto'

解决方案 #

1. 强化系统提示词约束 #

在系统提示词中明确强硬地要求AI必须校验参数完整性，禁止编造参数：

你需要协助用户填写XYZ报告，仅能调用`fill_form`函数完成。调用前必须严格检查对话历史，只有当所有必填参数（name、address、year_end、zip）都已从用户处获取时，才能触发函数调用；若有任何一个参数缺失，必须向用户询问对应的信息，绝对不能编造参数值。

2. 手动校验参数后再触发函数调用 #

在调用OpenAI API之前，自行编写逻辑检查对话历史中是否已收集到所有必填参数：

• 如果存在缺失参数，直接生成询问用户的消息，添加到对话列表后调用API并设置function_call='none'，让AI返回询问内容；
• 如果所有参数齐全，再调用API并指定function_call={"name": "fill_form"}强制触发函数调用。

示例代码逻辑：

# 定义必填参数列表
required_params = ["name", "address", "year_end", "zip"]
params_collected = {}

# 遍历对话历史提取用户提供的参数（需根据实际对话内容做针对性解析）
for msg in prompts:
    if msg["role"] == "user":
        # 示例：从用户输入中提取参数，实际需根据对话内容调整解析逻辑
        # 比如用户说"我叫李四，住在Oak Avenue 456"，则提取name="李四"，address="Oak Avenue 456"
        # 这里仅为逻辑示例，实际需实现具体的参数提取逻辑
        pass

# 找出缺失的参数
missing_params = [param for param in required_params if param not in params_collected]

if missing_params:
    # 生成询问用户的消息
    prompt_msg = f"请补充以下必填信息：{', '.join(missing_params)}"
    prompts.append({"role": "assistant", "content": prompt_msg})
    # 调用API返回询问内容，不触发函数调用
    completion = openai.ChatCompletion.create(
        model='gpt-4-0613',
        messages=prompts,
        functions=[
            # 原函数定义
            {
                "name": "fill_form",
                "description": "Helps the user create an XYZ Report",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "name": {"type": "string", "description": "the full name of the person issuing this report"},
                        "zip": {"type": "string", "description": "the 5 digit zip code of the address"},
                        "address": {"type": "string", "description": "the street address, only the street and not the city, state or zip"},
                        "year_end": {"type": "string", "description": "the full four digit year of the fiscal year"},
                    },
                    "required": ["name", "address", "year_end", "zip"]
                }
            }
        ],
        function_call='none'
    )
else:
    # 参数齐全，强制调用函数
    completion = openai.ChatCompletion.create(
        model='gpt-4-0613',
        messages=prompts,
        functions=[
            # 原函数定义
            {
                "name": "fill_form",
                "description": "Helps the user create an XYZ Report",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "name": {"type": "string", "description": "the full name of the person issuing this report"},
                        "zip": {"type": "string", "description": "the 5 digit zip code of the address"},
                        "address": {"type": "string", "description": "the street address, only the street and not the city, state or zip"},
                        "year_end": {"type": "string", "description": "the full four digit year of the fiscal year"},
                    },
                    "required": ["name", "address", "year_end", "zip"]
                }
            }
        ],
        function_call={"name": "fill_form"}
    )

3. 启用函数调用的强制校验模式 #

通过设置function_call为指定函数，并在系统提示词中反复强调参数必须来自用户输入，结合API的函数调用校验机制，减少AI编造参数的概率。不过该方法效果可能有限，建议配合前两种方法使用。

内容的提问来源于stack exchange，提问作者LJF 2525