(Case) Soma Health · 2025
A clinical-grade RAG assistant for 14k providers.
Client · Soma HealthYear · 2025Discipline · AI · ML
The brief
Soma's clinicians wanted guideline-grounded answers in under 2 seconds, with citations a regulator would accept.
Our answer
A retrieval pipeline over 1.2M clinical documents, hybrid BM25+dense, with a fine-tuned reranker, eval harness, and a hallucination-rate gate in CI.
1.4s
Median latency
98.7%
Citation accuracy
0.6%
Hallucination rate
Stack
Built with.
- Python
- pgvector
- Pinecone
- vLLM
- Anthropic
Next case
Ember Commerce
Headless storefront, agentic merchandising.