Note search
Free-text query is encoded with the same fine-tuned model that produced the corpus, then ranked by cosine similarity over the FAISS index.
Evidence Q&A
Top-k retrieved notes are passed to an LLM as evidence. Evidence mode lets the model synthesize patterns across cases with [Note N] citations; strict mode refuses unless the literal answer is in the retrieved text (for grounded-retrieval benchmarks).
Cohort discovery
Find clinically similar patients in two ways. By patient uses each patient's mean-pooled embedding to retrieve neighbors of a seed patient, reporting ICD-9 chapter Jaccard overlap. By description encodes a free-text clinical phenotype with the same fine-tuned model and ranks patients by cosine similarity.
Patient trajectory
Per-patient sequence of anchor notes ordered chronologically. The velocity plot shows L2 distance between consecutive note embeddings — spikes are candidate inflection points (acuity changes, transfers). The PCA plot is a 2-D projection of the patient's note embeddings, colored by note order.