phuonguyen/medical-rag-with-gnn

title: Medical RAG with GNN emoji: 🏥 colorFrom: blue colorTo: green sdk: docker app_port: 8501 tags:

• —streamlit
• —gnn
• —rag
• —medical pinned: false license: mit ---

🏥 Medical RAG with GNN

A modern, graph-powered RAG pipeline for medical QA.

This project combines a knowledge graph, GNN-based structural embeddings, and a hybrid retriever to deliver accurate medical answers. All components are optimized for CPU and Hugging Face Spaces.

Quick Navigation

• —[System Diagram](#system-diagram)
• —[Artifacts & Data](#artifacts--data)
• —[Graph & GNN Details](#graph--gnn-details)
• —[Retriever Logic](#retriever-logic)
• —[App & Usage](#app--usage)
• —[Deployment Notes](#deployment-notes)
• —[Directory Layout](#directory-layout)
• —[Citations](#citations)

System Diagram

MedQA Subset
	│
	▼
PropertyGraphStore (LlamaIndex)
	│
	▼
GCN (PyTorch Geometric)
	│
	▼
Hybrid Retriever (Semantic + Structural)
	│
	▼
Streamlit UI

Artifacts & Data

• —Graph: src/graph_storage_500/ (JSON)
• —Embeddings: src/pyg_data.pt, src/structural_embeddings.pt
• —Node Map: src/node_id_map.pt
• —Subset: MedQA (200 docs)

Graph & GNN Details

• —Graph Construction: Entities/relations from medical Q&A via LlamaIndex.
• —GCN Model: 2-layer, 384-dim, self-supervised link prediction. See src/rag_gnn_structural_embeddings.py.
• —Artifacts: PyG Data, GNN embeddings, node map.

Retriever Logic

File: src/dual_scoring_retriever.py

• —Semantic Score: Cosine similarity (query ↔ node text embedding)
• —Structural Score: Cosine similarity (query ↔ GNN embedding)
• —Hybrid: Weighted sum (α slider in UI)
• —Returns: Top-5 nodes for context

App & Usage

• —Entry: app.py (root)
• —UI: Streamlit (question input, α slider, context expander)
• —Artifacts: Loaded at startup

How to run:

1. 1.Launch the Space (first run may download LLM weights)
2. 2.Enter a medical question
3. 3.Adjust α for retrieval balance
4. 4.Expand context to inspect retrieved nodes

Deployment Notes

• —Docker: Uses sdk: docker for Hugging Face Space
• —RAM/CPU: 16GB RAM, CPU-only
• —Artifacts: All required files are committed

Directory Layout

.
├── Dockerfile
├── README.md
├── app.py
├── requirements.txt
└── src/
	├── dual_scoring_retriever.py
	├── rag_gnn_structural_embeddings.py
	├── node_id_map.pt
	├── pyg_data.pt
	├── structural_embeddings.pt
	└── graph_storage_500/
		├── property_graph_store.json
		├── index_store.json
		├── image__vector_store.json
		├── graph_store.json
		├── docstore.json
		└── default__vector_store.json

Citations

• —LlamaIndex
• —PyTorch Geometric
• —MedQA (Jin et al., 2020)
• —Streamlit