KG-MMRAG

Knowledge Graph–Augmented Multimodal RAG for Medical Imaging

Two retrieval paths feed MedGemma-4B: a dense vector index over SLAKE QA pairs (CLIP + FAISS, top-3) and a clinical knowledge graph (networkx, 1-hop, ≤8 triples). Toggle each path to reproduce the ablation from the report.

⚠️ Research prototype from IDEAS-TIH, ISI Kolkata. Not a medical device. Not for clinical use.

Medical image

Question

Vector retrieval

KG retrieval

Untick both for the no-retrieval baseline. The gap between that and both-on is exactly what the ablation measures.

Answer

Examples from the SLAKE test split

Code, full metrics, and significance tests: github.com/nathanpereira1234/Knowledge-Graph-RAG-Medical-Imaging