A local knowledge-graph + Graph RAG agent that helps US veterans file accurate, well-rated VA disability claims.
The agent ingests 38 CFR Part 4 (the VA rating schedule) into a Neo4j knowledge graph, asks veterans the right questions based on their military job history and exposure, and produces copy-paste-ready claim narratives in CFR vocabulary — one per claimable condition.
This is a learning project. It is local-only, not deployed, not commercial. The primary goal is hands-on Graph RAG learning; the secondary (and load-bearing) goal is a tool the author would have wanted for his dad.
v1 in implementation. See issue #1 for the PRD; implementation issues are tracked at #2–#13. Domain glossary lives in CONTEXT.md.
- Python 3.12 +
uv - Neo4j 5.x (Docker)
- LangChain 1.x + LangGraph
- Jupyter notebook as the primary dev surface
- 38 CFR §4.71a (musculoskeletal) + §4.1–§4.31 (general provisions)
- VA Duty MOS Noise Exposure Listing (all branches, all codes)
- Hand-curated musculoskeletal MOS risk overlay
- Hybrid vector + graph retrieval (Pattern 3 GraphRAG)
- LangGraph chat orchestrator that elicits symptoms, matches diagnostic codes, drafts CFR-language narratives
- Three vet personas as regression eval fixtures
Deployment, multi-user, web UI, browser extension, other body systems, scraped advisory content. All on the v2+ roadmap.