The graph is the product, and it does not arrive with the package — you stand up Neo4j, run the extraction against the documentation, and import the triples before the three tools return anything. Budget for that first; afterwards the payoff is a query shape ordinary documentation search cannot give you, which is "what do I need to understand before this makes sense".
An MCP server over a knowledge graph built from Unreal Engine's official documentation. Rather than returning matching pages, it returns relationships: which concepts a given concept contains, supports or depends on, each with the sentence of context that established the link. Queries take a Chinese and an English term together.
- `search_concept_relations` — related concepts for one term, each with a predicate, context and direction; takes `concept` as a `cn` / `en` pair, `limit` defaults to 20
- `search_concepts` — fuzzy name lookup when you remember the idea but not what it is called; `limit` defaults to 10
- `get_all_concepts` — the concept list ranked by how many relationships each has, with incoming, outgoing and total counts, so the core concepts sort to the top; `limit` defaults to 100
- Relationships stored as subject-predicate-object triples with a confidence score, extracted from the docs by DeepSeek v3
A running Neo4j instance — the repository ships a docker-compose file for it — and Node.js >= 18.0.0. Run the server with `npx -y unreal-engine-knowledge-graph-mcp` and pass `NEO4J_URI`, `NEO4J_USER` and `NEO4J_PASSWORD` in the env block. The database starts empty: populate it with `npm run extract-triplets` then `npm run import-to-neo4j`, and extraction needs a `DEEPSEEK_API_KEY`. `npm run extract-triplets:test-mode` processes a single document if you just want to see it work. Package `unreal-engine-knowledge-graph-mcp` at 0.1.1, MIT.
One command plus a key — npx -y unreal-engine-knowledge-graph-mcp, then supply credentials