RAG over clinical notes with structured entity extraction.
Status: Planned: Phase 4 of build plan
A demo project showing how to build domain-aware RAG for healthcare. Per-note chunking, metadata-filtered retrieval, and parallel structured extraction of medications, diagnoses, and vitals.
Live demo: coming soon Built on: genai-rag-template
Two design choices make this different:
Per-note chunking with rich metadata. Clinical notes are short and self-contained. Generic chunking fragments them across token windows. This assistant treats each note as a coherent unit and stores {patient_id, note_type, author_role, date, department} as metadata for filtered retrieval.
Dual output: chat + structured extraction. A chat answer goes on the left. A second LLM call runs structured extraction in parallel and pulls entities into JSON, displayed on the right. You get both the conversational answer and the structured data — useful for downstream pipelines.
- Synthetic clinical notes generated for demo purposes (included)
- MIMIC-IV demo dataset (requires CITI training; instructions in
data/README.md)
Next.js, FastAPI, pgvector, Claude Sonnet. See the template repo for the full architecture.
- Synthetic note generation script
- Per-note chunking with metadata
- Chat UI with citation links
- Structured extraction sidebar
- MIMIC-IV loader
- Public demo deployed
This is a demo project. Do not use real patient data.
The hosted demo accepts only synthetic clinical notes. The system is not HIPAA-compliant, does not offer a Business Associate Agreement, and has not been audited for use with Protected Health Information.
Any text submitted to the demo:
- Is sent to third-party LLM APIs (Anthropic) for processing
- May be retained in application logs for debugging
- Is not encrypted at rest with healthcare-grade controls
By using the demo, you affirm that any text submitted contains no real patient information, no PHI as defined under HIPAA, and no data requiring a BAA to process. The synthetic notes included in the repo are safe to experiment with.
MIT