Builder · Live on a free-tier host · public · 2026
Speculative RAG
You ask a question out loud in one of four Indic languages and get an answer grounded in the documents, in under 200 ms.
- Retrieval
- ~25 mslive, on a free-tier 512 MB host, inside a 200 ms answer budget
How it works
The fast path turns speech into text with Sarvam, retrieves with BM25 and Qdrant, and pulls out the answer. If that isn't good enough, a slower background step checks whether the answer is grounded and stitches passages together. It runs live on a free-tier host with 512 MB of memory.
- LinkCode ↗