Literature Search with GraphRAG (Qdrant + Neo4j)

A hands-on workshop that builds a progressively more powerful biomedical literature search system — from a vanilla LLM to full GraphRAG.

Setup

cp .env.example .env        # Add your OpenAI API key
docker compose up -d         # Start Qdrant + Neo4j
pip install -r requirements.txt

Notebooks

#	Stage	What breaks	What fixes it
00	Setup	—	Verify connections, preview data
01	Vanilla LLM	Hallucinated citations	—
02	Vector RAG	Semantic drift (wrong papers)	Dense vector search + Qdrant
03	Hybrid Search	Missing keyword precision	BM25 + dense + RRF fusion
04	GraphRAG	Can't find foundational/connected papers	Neo4j citation graph enrichment

Run them in order. Each one exposes a limitation that the next one solves.

Architecture (final)

Question → [Qdrant: Dense + BM25] → Retrieved Papers
                                          ↓
                                    [Neo4j: Citations, MeSH, Genes]
                                          ↓
                                    [LLM with full context] → Grounded Answer