whoknowsasaint/reposcope

Reposcope is a tool that lets you ask plain-English questions about any GitHub repository's code and get back answers with exact file paths and line numbers. It uses a technique called RAG (Retrieval-Augmented Generation), which means it first indexes the codebase into a searchable form, then uses that index to retrieve relevant code snippets before generating an answer. The indexing step clones the repository and splits the code into chunks using AST-based (Abstract Syntax Tree) chunking via tree-sitter, which splits code at meaningful boundaries like functions, classes, and methods rather than arbitrary character counts. Those chunks are turned into vector embeddings, numerical representations that capture meaning, and stored locally in ChromaDB, a local vector database. When you ask a question, the most relevant chunks are retrieved and passed to an LLM (Groq running Llama 3.3 70B) to generate the answer. Conversation history is preserved across questions in a local SQLite database, so you can ask follow-up questions that reference earlier context. You can use it either through a command-line interface or through an optional web interface built with Next.js and Tailwind CSS. The backend is FastAPI, the CLI uses Click and Rich, and embeddings are generated using Jina AI's jina-embeddings-v3 model. Everything, cloned repos, embeddings, and conversation history, is stored locally in a folder in your home directory. The project is written in TypeScript for the frontend and Python for the backend, and is licensed MIT.