Meta-programming with the new function calling feature

Community
,
1

I tried some kind of meta-programming with the new feature of function calling.
It is funny to see, that GPT-3 can generate a function which I define in the prompt and call the function calling API.

This is the function call in Json:

{
    "name": "evaluate_Code",
    "description": "A function to evaluate code to ensure correctness.",
    "parameters": {
        "type": "object",
        "properties": {
          "functionName": {
            "type": "string",
            "description": "The name of the created function to evaluate, e.g. factorial, fibonacci, print, etc"
          },
          "args": {
            "type": "string",
            "description": "The parameters to pass to the function for evaluation, e.g. n: int, x: int"
          },
          "body": {
            "type": "string",
            "description": "The entire function body to evaluate"
          }
        },
        "required": ["functionName", "args", "body"]
    }
}

Here is the prompt I use:

Show me first a python function to calculate the area of a circle. After then call this function for evaluation.

The response has both, a content and the function_call:

content:

Sure! Here’s a Python function to calculate the area of a circle:

import math

def calculate_area(radius):
    area = math.pi * radius ** 2
    return area

Now, let me call this function for evaluation.

And then this is the function_call to my backend:

{
   'name': 'evaluate_Code', 
   'arguments': '{\"functionName": "calculate_area", \"args": "radius: float", \"body": "import math\\n\\narea = math.pi * radius ** 2\\nreturn area"\}'}}

You see, it does exactly what I want.
I used the gpt-3.5-turbo-0613 with a temperature of 0 (which I think is very important)

What are your use cases and what benefits can this have?

Some will ask "For what is this useful?.

I think, this is some kind of universal API endpoint, which can execute arbitrarily code.

6 Likes
2

Let us find these people and slap them, I say.

1 Like
3

To save on tokens, the following also works.

“Return a python function to calculate the area of a circle, in JSON format.”

The output:

Here’s a Python function to calculate the area of a circle:

import math
def calculate_circle_area(radius):
    area = math.pi * radius**2
    return area

And here’s the function converted to JSON format:

{
  "function": "calculate_circle_area",
  "code": "import math\n\ndef calculate_circle_area(radius):\n    area = math.pi * radius**2\n    return area"
}

The JSON format includes the function name (calculate_circle_area) and the corresponding Python code for the function.

4

is the entire code count as a token
i dont know how fuction api work

is it like prompt eng?