clinical-graphrag-gxp

Drug-label question answering for medical-information and pharmacovigilance staff: GraphRAG on Neo4j with guardrails and human review, a PySpark + Delta Lake pipeline, MLflow lineage and approvals, Langfuse tracing, a six-dimension evaluation and a generated GxP-style validation package. Code: GitHub.

Public openFDA data. Every number below is read from results/*.json written by the evaluation code (NVIDIA L4, local models at temperature 0). GxP-style demo, not a validated system. Pipeline ran on local PySpark + Delta Lake, not on a Databricks workspace.

Demo

Every command in the video ran live against the services while it was rendered.

Three systems on the same questions

All metrics as a table (all runs, with n)

PQ acceptance criteria: first run (v1, blocked) and CAPA re-run (v3, released)

Frozen in config/acceptance_criteria.yaml before the PQ run.

Release gate decisions

Real traced answers from the released PQ run (system C, v3)

Each answer has a Langfuse trace; the bar shows the span durations in order (guard, route, query or retrieval, generation, citation check).

Root cause analyses and change control

Traceability matrix

Show the matrix (URS -> risks -> OQ tests -> PQ criteria -> release)

Data pipeline

Limitations

No Databricks workspace: same PySpark / Delta code run locally; the Asset Bundle was never deployed.
Gold answers come from structured openFDA fields; open clinical reasoning is not measured.
The measured human review uses an oracle reviewer, so accuracy after review is an upper bound.
Identities are not authenticated; approvals are Part 11-style records, not compliant e-signatures.
7-8B local models on one GPU; latency is for this host only.