Clinical Knowledge Graph RAG

The problem

Researchers needed to ask free-text questions across clinical trial, drug and disease data — but the answers depend on many-to-many drug–disease–dosage relationships that chunk-based retrieval flattens and loses.

Why it needed solving

A conventional RAG pipeline retrieves passages that look similar to the question. That works for definitional lookups and fails for relationship questions, where the answer is not contained in any single chunk but in how entities connect. Establishing that limit empirically was what justified the redesign.

The approach

Build the conventional pipeline first, evaluate it honestly, identify exactly where it broke, then model the domain as a graph and translate natural-language questions directly into graph queries.

Architecture

• Ingestion — clinical trial, drug and disease sources, chunked with metadata tagging so retrieval can filter as well as match.
• Retrieval v1 — hybrid retrieval combining semantic vector search with BM25 lexical search.
• Knowledge graph — entities and relationships modeled in Neo4j, capturing drug–disease–dosage links explicitly rather than implicitly.
• Scale evaluation — AWS Neptune assessed as an alternative for larger-scale ingestion.
• Query translation — agent-based layers converting free-text clinical questions into Cypher (Neo4j) and Gremlin (Neptune).
• Serving — a research-facing Q&A chatbot over the retrieval layer.

Implementation

• Built the metadata-tagged chunking strategy so retrieval could constrain by trial, drug or disease rather than relying on embedding similarity alone.
• Combined semantic and BM25 retrieval to cover both paraphrased and exact-term clinical queries.
• Evaluated the pipeline against relationship-heavy questions and documented where it could not represent many-to-many links.
• Modeled entities and relationships in Neo4j and contributed to the Graph RAG redesign built on that model.
• Built natural-language-to-Cypher and natural-language-to-Gremlin translation workflows so researchers never write query syntax.
• Iterated on embedding and metadata strategies to improve entity-relationship linking.

AI/ML components

Challenges & how they were handled

Results

What comes next

• Expand relationship coverage beyond drug–disease–dosage to additional clinical entity types.
• Formalize retrieval evaluation for graph traversal answers, not just passage relevance.