Databricks knowledge graph
Except it isn’t, not really.
Maybe I should stop mentioning companies in the semantic space without anything semantic, but I hope it helps to keep differentiating what is and what isn’t.
What’s actually happening: Unity Catalog tables get a graph schema bolted on top (via tools like PuppyGraph), or data gets piped out to Neo4j and back. That’s graph access on top of a lakehouse, not a knowledge graph.
A knowledge graph isn’t a query pattern. It’s a semantic layer with entities, relationships, and an ontology that encodes meaning, not just foreign keys with a new syntax. Calling a SQL table a node doesn’t make your data understand itself.
This matters because the GraphRAG hype is pushing every data platform to slap “knowledge graph” on their roadmap, when most of what’s shipping is graph-shaped querying over relational data. Which is useful, but not the same thing. The actual hard part (entity resolution, ontology design, relationship extraction from unstructured sources) is exactly the part being skipped.
If you’re building GraphRAG on top of your lakehouse and skipping the semantic modeling step, you’re not building a knowledge graph. You’re building a relational database with extra syntax.