Codexa

Codexa is a multi-agent AI system that lets you explore any public GitHub repository through natural language. Paste a repo URL and ask questions and get back grounded answers with exact file and line citations, eliminating guesswork when navigating unfamiliar codebases. It combines semantic RAG with a LangGraph orchestrator, where code is parsed using tree-sitter, embedded with Vertex AI, and stored in a FAISS vector index. A lexical reranking step refines retrieval for a hybrid search pipeline, while answers stream token by token via Groq with sub-2-second time to first token.

The agent graph includes specialized roles covering retrieval, generation, memory, planning, repo analysis, and validation, enabling both a low-latency fast path and a deeper multi-step reasoning mode. Conversational memory ensures follow-up questions resolve correctly across turns. The backend is built on FastAPI, the frontend on Next.js with Tailwind CSS, and the full system is containerized with Docker Compose and deployed on GCP Compute Engine behind Caddy for automatic HTTPS. A live evaluation dashboard powered by Prometheus tracks latency, agent utilization, and per-repo analytics, with CI/CD via GitHub Actions handling linting, testing, and builds on every push.