GitHub Code Reader

A web application that enables developers to search, browse, and chat with their GitHub repositories using Retrieval-Augmented Generation (RAG). By combining vector-based semantic search with LLMs, users can ask questions about codebases, get relevant code snippets, and locate specific functionality across multiple files.

Challenges

•

Crawling and parsing GitHub file trees dynamically while staying within API rate limits and file size constraints

•

Designing a custom boundary-aware chunker to split source code by logic boundaries (classes/functions) rather than arbitrary characters

•

Optimizing query latency for real-time semantic search and LLM response streaming over large codebases

Solutions

•

Implemented recursive branch crawling using @octokit/rest, filtering files under 100KB to reduce resource usage

•

Developed a regex-driven chunker in lib/chunker.ts to partition code files along function boundaries, falling back to overlapping sliding windows

•

Utilized Cohere embeddings with Pinecone vector search for low-latency retrieval, and streamed Groq API llama-3.3 completions directly to the frontend