gnaw

gnaw's Vision

The Problem 🚩

Large Language Models (LLMs) have revolutionized the way we interact with code. However, they still face significant challenges with code generation:

  • Planning and Reasoning: LLMs lack the ability to plan and reason, which is crucial for tasks like code generation, refactoring, and debugging. They often struggle to get the big picture and are short sighted.
  • Context size: LLMs have a limited context window, which restricts their ability to analyze and understand large codebases.
  • Hallucination: LLMs can generate code that appears correct but is actually incorrect or nonsensical. This phenomenon, known as hallucination, occurs when the model lacks sufficient context or understanding of the codebase.

This is where gnaw comes in.

The Solution ✅

We believe that planning and reasoning can be achieved by human or AI agents with scaffolding techniques. These agents need to gather a high quality context of the codebase that is filtered, structured, and formatted for the task at hand.

The thumb rule would be:

This is practically difficult to achieve, especially for large codebases. However, gnaw is a simple tool that can help developers and AI agents ingest codebases more effectively.

It automates the process of traversing a codebase, filtering files, and formatting them into structured prompts that LLMs can understand. By doing so, it helps to mitigate the challenges of planning, reasoning, and hallucination.

Architecture ⛩️

gnaw is designed in a modular way, allowing for easy integration into various workflows. It can be used as a core library, a command line / terminal interface (CLI/TUI), through Python bindings, or — in the future — as a REST service for browser-extension integration.

Core

gnaw-core is a code ingestion library that streamlines the process of creating LLM prompts for code analysis, generation, and other tasks. It works by traversing directories, building a tree structure, and gathering information about each file. The core library can easily be integrated into other applications.

CLI / TUI

The gnaw command line interface was designed for humans to generate prompts directly from your codebase. The generated prompt is automatically copied to your clipboard and can also be saved to an output file. An interactive TUI lets you toggle files in and out of the context and watch the token budget update live. Prompt generation is customizable using Handlebars templates.

Python

The Python bindings expose the core library for AI agents or automation scripts that want to interact with codebases seamlessly.

REST

A REST interface (axum) is planned, so that browser extensions and other clients can request well-structured context over HTTP.