LiteralAI compiler for Python

What is literal ai coding? It's the opposite of agentic AI enabled IDEs like Cursor, where your prompts modify your entire codebase and are then quickly forgotten, leading to write-only throw-away code. Literal ai is a way to store your prompts inside your project, and check them into git just like the rest of your code. It works more like a compiler - compiling your prompts into source code.

How it works

Any function that only has a docstring and/or initial comments, will be generated when literalai is run on your project, and any function that was generated this way, will be updated if you change the signature, docstring and/or comments. The LLM prompt will include the function signature, docstring and any initial comments. Literal ai does not store any metadata or state outside of your sourcecode, and the only state it stores is a hash of the signature/docstring/comments, as an extra comment inside functions it generate.

Example

You write

def add_two(a, b):
    """Add two numbers together, return the result"""

def manual(x):
    "A manual function"
    return x + 1

and run literalai .

The file will then be updated to contain

def add_two(a, b):
    """Add two numbers together, return the result"""
    # LITERALAI: {"prompt": "somehash"}
    result = a + b
    return result

def manual(x):
    "A manual function"
    return x + 1

Notice how the manual function stays untouched. If you where to change the docstring of add_two and rerun literalai ., its body would be regenerated using the new docstring.

What about classes?

Any class that only has a docstring and/or initial comments, will be generated when literalai is run on your project, and any class that was generated this way, will be updated if you change the signature, docstring and/or comments.

What it means to update a class is that method signatures and docstrings will be generated matching the docstring of the class. These will then be generated as functions as per above. Any methods previously generated this way and whos' docstrings or comments haven't been manually changed, will be replaced by the new ones.

Methods that have manually written or changed signatures, docstrings or initial comments won't be touched / overwritten.

Models, system prompts and other configuration

Model names, system prompts / prompt templates and other configuration is read from yaml files literalai.yml inside your project. All such files found along the path from the project root to your source file are merged and used as the config for that file.

Inside the config there are top level blocks: base, FunctionDef and ClassDef. The two latter overrides the base section for functions and classes respectively.

Each block can set the following keys:

• model - a litellm compatible model string such as openai/gpt-4
• prompt - a system prompt template processed using jinja2 with access to the variable signature (the function or class signature) and all of the config.

Note: Changes to the config will lead to code regeneration for affected code (e.g. base and FunctionDef regenerates functions).

Example literalai.yml:

base:
  model: "openai/gpt-4"

FunctionDef:
  prompt: |
    Generate the python source code for a function with the following
    signature, docstring, and initial comments.

    {{signature}}

    # IMPORTANT
     * Write the full function implementation.
     * Provide only valid python for a single function as output.
     * Do NOT add any initial description, argument or similar

ClassDef:
  prompt: |
    Below is the python source code for a class and some of its methods
    (without implementations). Given the docstring and initial comments of
    the class, define any missing method signatures and provide their
    docstrings.

    {{signature}}

    # IMPORTANT
     * Write the full class specification.
     * Provide only valid skeleton python for a single class as output.
     * Do NOT add any initial description or similar

Installation

pip install literalai-python