The function call is not good

Prompting
, ,
1

Hi, I’m new user
I using model gpt-3.5-tubro-0613 with funtions call
I have 4 func and confrim funtions

{
    "name": "confirm_function_submission",
    "description": "MUST use to confirm whether to submit request from function.",
    "parameters": {
        "type": "object",
        "properties": {
            "confirmation": {
                "type": "string",
                "description": "Either Yes or No, based on response to 'Do you want to submit (yes) or cancel (no) your request?'",
            },
            "previous_function": {
                "type": "string",
                "description": "The name of the previous function they are confirming.",
            },
        },
        "required": ["confirmation", "previous_function"]
    }
};

The prompt exam

[
    {
        "role": "system",
        "content": "..... bla bal"
    },
    {
        "role": "user",
        "content": "..... bla bal"
    },
    {
        "role": "system",
        "content": "the {funtions_name} requested and waiting on user confirmation. Call the confirm_function_submission function after they confirm whether to submit the request."
    },
    {
        "role": "assistant",
        "content": "Do you want to submit (yes) or cancel (no) your request?"
    },
    {
        "role": "user",
        "content": "Yes"
    }
]

However, it does not really bring into question functions when I reply “yes” or “no”
Does anyone have any suggestions for this problem?

2

It appears you have a perception of how the AI should work for function-calling that is quite different from how AI models have been trained.

Now:

  • User asks AI a question beyond its typical abilities;
  • AI calls a function that can enhance its abilities.

Proposed:

  • User asks AI a question beyond its typical abilities;
  • AI calls a function that can ask user a question.

I think you might do better in the perceived goal by just placing in your instructions a prohibition on calling API tool functions unless the user has confirmed and made clear their desire for AI to use such tools, which shall not be called autonomously. Explanation of what AI wants to do first. That gets you confirmation asked by AI in its language.

You can also consider programmatically prompting for confirmation of the function execution in your UI when it is emitted from the AI.

3

I see what you mean. In fact, I want to manage more logic by having GPTChat call that location. I attempted to change the functions description, but it is not very good.

4

I have the same problem. Does anyone have any ideas for this issue?