Knowledge graphs in practice
Short answers to the questions that come up before and during a knowledge graph project: whether you need one, how to model it, how to test it and what it costs. Each guide answers one question, shows an example from a graph I built, and says when the answer is "don't".
- GraphRAG vs vector RAG: when a graph helps
What 2025-2026 benchmarks say about GraphRAG vs vector RAG: where graphs win, where plain RAG ties, what graphs cost, and how to test it on your own data.
- When not to build a knowledge graph
Five signs you can skip a knowledge graph, what to use instead (SQL, search, vector RAG), and the three conditions under which a graph earns its cost.
- How to test a knowledge graph
Test a knowledge graph with competency questions, zero-row invariants and count tripwires in Cypher, run on every load. With a worked 10-K example.
- Reuse, extend or build an ontology: how to decide
When to reuse a standard ontology such as SNOMED CT, FIBO or EPPO, when to take a slice and extend it, and when to build your own.
- What a knowledge graph costs
What building, running and maintaining a knowledge graph costs: where the money goes, what makes a graph cheap or expensive, the costs people forget, and how to estimate your own.
- How to make AI answers auditable with provenance
An auditable AI answer names the source of each fact and how far it can be trusted. How to model provenance and fact status in Neo4j, and what it misses.