A large children's hospital has put semantic search to work on its entire patient record archive — and the results make keyword search look embarrassing.
Researchers deployed a system indexing 166 million clinical notes from 1.68 million patients, generating 484 million embedding vectors using a small instruction-tuned model called qwen3-embedding-0.6B. The whole thing runs with sub-second query latency and costs roughly $4,000 a month to operate. The team evaluated it against a physician-authored benchmark, where it hit 94.6% accuracy using 300-token text chunks — then tested it on real clinical work.
The gap between semantic search and the standard alternative — ICD-10 diagnosis codes — is where this gets interesting. When hunting for patients with molecularly confirmed genetic diseases, semantic search recovered 98% of them; diagnosis codes found at most 75%. That 23-point gap matters because ICD-10 codes are how nearly every hospital currently builds research cohorts and tracks populations. If semantic search can reliably outperform billing codes at finding sick patients, the implications for clinical research enrollment alone are significant.
Chart review speed also improved sharply: across three abstraction tasks, the system cut time-to-completion by 24% to 89% compared with manual review. The authors are careful to note the system operates within a HIPAA-compliant governance framework — a requirement that has historically been the silent killer of promising hospital AI pilots. Whether other institutions can replicate a $4,000-a-month price tag at comparable scale remains the open question; this was one well-resourced children's hospital, not a sprawling multi-site health system.