A retrieval assistant answering ~50,000 people a day
Built a RAG assistant that gives fast, grounded answers from a million-document corpus — without hallucinating its way into a problem.
The challenge
A healthcare organization needed to answer a high volume of questions on demand from a corpus of more than a million documents. Keyword search was slow and unreliable, and in a regulated, patient-facing setting a confidently wrong answer wasn't acceptable.
What we did
We built a document-processing and vectorization pipeline, then combined vector search with keyword matching for hybrid retrieval that stayed grounded in the source material. Answers streamed to a React front end with a server-sent-events fallback, and Redis caching plus query tuning brought latency down hard.
The outcome
The assistant serves roughly 50,000 people a day. Median response time dropped from 3 seconds to half a second, and answer accuracy improved about 40% over the keyword-only baseline it replaced.