Quansight / pseudocode

PseudoCode

There is a blog post about this project.

pseudocode is a python module which allows users to describe the functions they want via type annotations and docstrings without writing the actual code. It empowers users to iterate with large language models to generate functions which satisfy the user along with defined tests.

pseudocode enforces the defined tests and provides automated feedback to the LLM when attempting to run these tests by reporting exceptions. There is a built in review cycle for the user to provide feedback on the generated code to help guide the LLM.

from pseudocode import pseudo_function

@pseudo_function(review=True)
def my_sum_func(a: int, b: int, c: int) -> str:
    """A function which adds all arguments together and multiplies by two

    Examples
    --------
    >>> my_sum_func(1, 2, 3)
    12
    """
    pass

@pseudo_function
def get_total_issues(repository: str) -> int:
    """Get the total issues of given github repository

    Examples
    --------
    >>> get_total_issues("conda/conda")
    """
    pass


# LLM generated functions!
print(my_sum_func(1, 4, 5)) 
print(get_total_issues('conda/conda-libmamba-solver'))

This work uses OpenAIs gpt-3.5-turbo at the moment. Because of this it expects an environment variable OPENAI_API_KEY to be defined.

Vision

Much like how high level languages such as python are abstractions above lower level languages like C/C++. psuedocode aims to take advantages of LLMs to work with developers and enable them to create high level well defined interfaces. LLMs will fill in the details with iteration and feedback from the user.

Longer terms goals:

• psuedocode.pseudo_file function to generate non-python files e.g. docker-compose.yaml, Dockerfile, pyproject.toml, etc.
• caching of functions generated from pseudocode.pseudo_function
• richer testing utilities for defined interfaces

License

MIT