kyutai-labs / moshi-rag

What is moshi-rag?

MoshiRAG (aka moshi rag or moshirag) is a compact full-duplex speech-to-speech AI model that runs real-time voice conversations while asynchronously fetching external knowledge to boost factuality. Built on the Moshi voice model, it predicts when to trigger RAG retrieval—like for factual queries—injects references mid-response without pausing the chat, and delivers natural interactivity. Developers get PyTorch for experiments, Rust for production servers, plus a web UI demo with visualizations, transcripts, and retrieval panels; deploy via Docker Compose or Swarm.

Why is it gaining traction?

It stands out by marrying full-duplex voice (listen/speak simultaneously) with seamless asynchronous RAG, improving answer accuracy without latency spikes that plague other voice AIs. Users notice grounded responses in live chats, like fact-checked trivia during casual talk, via low-latency local LLMs (e.g., vLLM) or APIs. The web client shines with audio stats, pipe animations for retrieval flow, and exportable transcripts/videos—hooks for quick demos.

Who should use this?

Voice AI builders needing factual grounding in real-time apps, like interactive tutors or assistants handling queries beyond training data. Researchers tweaking speech models (e.g., Moshi variants) for duplex setups with RAG. Devs prototyping multilingual voice agents, especially those eyeing Rust efficiency or Hugging Face model integration.

Verdict

Promising for voice+RAG experiments, with solid docs, models on HF (CC-BY-4.0), and a runnable demo—but at 44 stars and 1.0% credibility, it's early-stage; expect tweaks for prod stability. Try the web UI if full-duplex factuality intrigues you.

(198 words)