A RAG agent for medical inquiries — accurate, cited, and refusal-safe.

A retrieval-grounded agent that answers clinician and patient questions from a compliance-bound knowledge base, with citations on every answer.

−72%

average response time on inbound inquiries

98.4%

answer-citation match rate on the eval set

policy-violating answers in 90 days of production traffic

Healthcare · Compliance

Medical Information Provider

Client

Medical Information Provider

Healthcare · Compliance

Headline metric

−72%

average response time on inbound inquiries

Deliverables

shipped to production

Stack

4+ tools

across the build

Challenge

A medical-information provider was answering a backlog of clinician and patient inquiries by hand. Every answer had to cite a source from a compliance-approved knowledge base. Throughput was the bottleneck; safety was non-negotiable.

Approach

Hybrid RAG (BM25 + embeddings) over the approved corpus, with strict citation requirements at generation time. The agent refuses cleanly when context is missing, and routes ambiguous queries to a clinician review queue. A continuous eval suite runs every prompt change against 1,200 reference answers.

Outcome

Average response time down 72%, citation-match rate at 98.4% on the eval set, and zero policy violations in 90 days of production. The clinician review queue now sees only the 6% of queries that genuinely need expert judgement.

“