AskRepo
A high-performance repository intelligence platform featuring a Next.js 16 dashboard and FastAPI backend that lets developers chat with any GitHub codebase in real time using Tree-Sitter AST chunking, two-stage vector retrieval with FlashRank ONNX reranking, and multi-LLM streaming failover.
About the Project
askRepo is a decoupled, high-performance codebase intelligence platform designed for deep repository navigation and real-time QA. Built with a Next.js 16 web dashboard and a Python FastAPI backend, it enables developers to index any public GitHub repository and converse with it seamlessly. Under the hood, a decoupled worker pipeline parses code at the Abstract Syntax Tree (AST) level using Tree-Sitter to extract precise logical symbols (functions, classes, and methods). Vector embeddings, generated locally via Ollama (embeddinggemma) or serverless GPU on Modal.com, are stored in PostgreSQL with pgvector. Query retrieval uses a two-stage pipeline combining vector similarity search with FlashRank ONNX cross-encoder reranking (~10ms) for high-precision context retrieval. Streaming responses feature automatic multi-provider LLM failover (Google Gemini, Groq / Llama 3.3, Mistral AI), supported by Redis session caching, mandatory file & symbol source citations, and an offline LLM-as-a-Judge RAG evaluation suite.