janekb04/py2gpt

Convert Python code into JSON consumable by OpenAI's function API.

py2gpt

Recently, OpenAI has released the GPT function API that allows models like GPT-4 to "invoke" functions on the user's behalf. The problem is that the format, in which these functions is specified, is JSON Schema, which is relatively niche and hard to work with. This repository addresses this issue. It converts regular Python function declarations into JSON consumable by the OpenAI API.

For example, this Python code:

def get_current_weather(location: str, format: Literal["fahrenheit", "celsius"]):
    """
    Get the current weather
    :param location: The city and state, e.g. San Francisco, CA
    :param format: The temperature unit to use. Infer this from the users location.
    """

Produces the following JSON that matches one of OpenAI's own examples:

{
    'name': 'get_current_weather',
    'description': 'Get the current weather',
    'parameters': {
        'type': 'object',
        'properties': {
            'location': {
                'description': 'The city and state, e.g. San Francisco, CA',
                'type': 'string'
            },
            'format': {
                'description': 'The temperature unit to use. Infer this from the
users location.',
                'type': 'string',
                'enum': ('fahrenheit', 'celsius')
            }
        }
    },
    'required': ['location', 'format']
}

Roadmap

• Support classes (treat them as objects with certain `"properties"
• Support methods (treat them as functions with an additional self parameter`)
• Support multiple functions
• Support Enums
• Support named tuples
• Make the code safe for local execution
• Refactor code and make locally executable

py2gpt's Issues

using it as a library

Hey there! Great little project you wrote! I find your implementation to be way better for handling functions and more focused and strict than langchain's one for example, it's great not having to use pydantic or some middle step

In fact I've used in my library, check it out: https://github.com/rogeriochaves/litechain/blob/main/litechain/contrib/utils/open_ai_functions.py

I've just made a small change:

     if (
-        len(func.body) != 1
+        len(func.body) == 0
         or not isinstance(func.body[0], ast.Expr)
         or not isinstance(func.body[0].value, ast.Str)
     ):
-        raise ValueError("The function body should only contain a docstring")
+        raise ValueError("The function body should contain a docstring")

This is because I want my functions to have actual execution bodies too, so LiteChain actually calls them directly. I've also removed the pyscript bits, as I'm just using it as a module.

It is working well, however I didn't want to maintain this inside LiteChain, do you have plans to publish it as a library? I can help providing updates as well