DocuMind - Ask Your Codebase Anything

Most documentation is either out of date, buried in the wrong folder, or just doesn't exist. DocuMind lets you talk to your codebase instead. You connect it to a GitHub repo or Notion workspace. It reads the files, chunks them, generates vector embeddings, and stores everything in Qdrant. When you ask a question, it searches semantically rather than by keyword, pulls the most relevant chunks, and passes them to an LLM that writes an actual answer with source links attached. The frontend is a chat interface built in React. Responses stream in real time over server-sent events, so you're not waiting for a spinner to dump a wall of text. Sources show up alongside the answer so you can check what the model is actually citing. The backend runs on FastAPI with a multi-agent setup. A query agent handles search and response generation. The embedding layer uses sentence-transformers, so you're not locked into one LLM provider for that part. The LLM connects through Ollama, which means you can swap models depending on what you need. The whole stack is self-hostable. You bring your own keys and run it on your own infra. Nothing routes through a third party except the model endpoint you configure. In practice it's useful for three things: onboarding new engineers (let them ask "how does auth work" instead of interrupting someone), debugging (ask where a particular error gets thrown and which file handles it), and navigating large codebases that nobody fully understands anymore. That last one comes up more than you'd expect.