A medical RAG system built on ontologies, papers, and patience

What it does

The system grounds clinical LLM outputs in a three-tier knowledge graph, linking private patient data (via MIMIC) to medium-level medical literature and bottom-level ontologies like UMLS inside Neo4j. Answers trace back to actual evidence instead of model weights.

The interesting bit

The “trinity” hierarchy is the core mechanic: patient records, medical papers, and dictionary concepts are explicitly connected, forcing retrieval to hop across graph layers before generating a response. The authors also provide a Docker demo that searches PubMed live to dodge the licensing wall around local textbook storage.

Key highlights

• Hierarchical evidence retrieval across patient records, medical literature, and UMLS ontologies via Neo4j.
• Built on the CAMEL multi-agent framework for pipeline construction.
• Ships with a Docker demo using live PubMed search to work around medical data licensing restrictions.
• Accepted at ACL 2025; paper and arXiv preprint available.

Caveats

• Reproducing the full stack requires buying copyrighted medical textbooks, applying for UMLS access, and licensing MIMIC IV data.
• The simplified example dataset only covers the top patient-record layer; medium and bottom-level alternatives are still a work in progress.
• The documented path requires OpenAI and Neo4j credentials, with no guidance on swapping either backend shown.

Verdict

A solid reference for medical AI researchers and clinicians who can clear the data-access hurdles. If you are looking for a drop-in RAG toy, the licensing and infrastructure overhead will send you elsewhere.