Production RAG platform with hybrid dense + BM25 retrieval, a hallucination-blocking confidence gate and multi-provider fallback.
A retrieval-augmented generation platform built for production use rather than demos. Documents are chunked, embedded and indexed for both dense (vector) and sparse (BM25) retrieval, then fused and re-ranked so the model only ever sees the most relevant evidence.
Answers stream back with inline citations. A confidence gate scores the retrieved evidence before generation and refuses to answer when support is weak — the model is never allowed to guess.
Provider outages are handled automatically: requests fall back across multiple LLM providers with identical prompts, so the product keeps working when one API degrades.
- Hybrid dense + BM25 retrieval with reciprocal-rank fusion and re-ranking
- Citation-grounded streaming answers
- Confidence gate that blocks low-evidence generations
- Automatic multi-provider LLM fallback
- Deployed with Docker, CI/CD and structured logging; runs in ~220 MB on a free tier