A voice assistant that answers from your own docs

Most voice demos are impressive right up until the assistant confidently makes something up. I wanted one that talks in real time but stays honest — answers pulled from real documents, not invented on the spot. That's VoiceRAG.

The loop

It's four pieces chained together: speech-to-text turns what you say into text, an LLM figures out what you mean, retrieval (RAG) pulls the relevant bits out of your own docs, and text-to-speech reads the answer back. Add tool-calling on top and it can go fetch things instead of guessing.

Why build it modular

The STT, LLM, and TTS models are all swappable. That matters because this space moves fast — a better model comes out every few weeks, and I didn't want to rebuild the whole thing every time. So you can drop in a new model, run the same tests, and actually compare them fairly. There's a small web interface for talking to it and poking at each part on its own.

Grounding is the whole point. A voice assistant that sounds great but invents answers is worse than useless — it's confidently wrong, out loud.