AI Ghostfunctions

Write the docstring, call the function, get the results.

Documentation: https://ai-ghostfunctions.readthedocs.io/

A ghostfunction leverages OpenAI GPT-4 to return sensible outputs from functions without defined logic. It packages the the function name, docstring, and signature into a prompt which is sent to OpenAI's GPT-4 API, and transparently handles coercing the api result into the expected return type.

Installation

pip install ai-ghostfunctions

Quickstart

To see it in action, save your OpenAI API Key to the env var OPENAI_API_KEY and run:

>>> import os
>>> from ai_ghostfunctions import ghostfunction

>>> assert os.getenv("OPENAI_API_KEY")

>>> @ghostfunction
>>> def sanitize_messy_string(messy_string: str) -> list[dict]:
>>>     """Return a list of dicts that contain the data from `messy_string`."""
>>>     pass

>>> sanitize_messy_string(messy_string="""name|age|nickname
John Brighton Bradford,  34,  J.B
     Grace B., "24", Grace""")

[{'name': 'John Brighton Bradford', 'age': 34, 'nickname': 'J.B'},
 {'name': 'Grace B.', 'age': 24, 'nickname': 'Grace'}]

###

>>> @ghostfunction
>>> def generate_random_words(n: int, startswith: str) -> list:
>>>     """Return a list of `n` random words that start with `startswith`."""
>>>     pass

>>> generate_random_words(n=4, startswith="goo")
['goofy', 'google', 'goose', 'goodness']

>>> generate_random_words(n=3, startswith="foot")
['football', 'footnote', 'footprint']

By default, a ghostfunction will dispatch a sensible prompt to OpenAI GPT-4 that includes the function name, the docstring, and function arguments, parse the result from OpenAI and return it as the result of the function.

Ghostfunctions will retry and send the data to gpt-3.5-turbo if it looks like the OPENAI_API_KEY does not have access to gpt-4.

Customizations

You can control the prompt:

>>> import os
>>> from ai_ghostfunctions import ghostfunction
>>> from ai_ghostfunctions.keywords import USER
>>> from ai_ghostfunctions.types import Message

>>> assert os.getenv("OPENAI_API_KEY")

>>> @ghostfunction(prompt_function=lambda f, **kwargs: [
>>>     Message(role=USER, content=f"tell me a slightly insulting joke about this function name: {f.__name__}.")
>>> ])
>>> def recursive_function_that_will_recurse():
>>>     """Recurse until you go crazy."""
>>>     pass

>>> recursive_function_that_will_recurse()
# 'Why did the programmer name his function "recursive_function_that_will_recurse"? Because he wanted to make absolutely sure that no one would confuse it for a function that actually does something useful.'

Heh. Not bad.

Prompts to gpt-4 and gpt-3.5-turbo are of type List[ai_ghostfunctions.types.Message].

See ghostfunctions.py for the default prompt.

License

AI Ghostfunctions is free and open source software.

Requirements

See pyproject.toml for a list of dependencies.

Contributing

Contributions are very welcome. To learn more about setting up a dev environment and contributing back to the project, see the Contributor Guide.

Issues

If you encounter any problems, please file an issue along with a detailed description.

Credits

This project was generated from a fork of @cjolowicz's Hypermodern Python Cookiecutter template.