OntoBricks turns Databricks Unity Catalog tables into a queryable knowledge graph. It runs as a Databricks App and uses existing workspace services: Unity Catalog for storage and metadata, a SQL Warehouse for execution, Model Serving for LLM automation, and Lakebase for the registry and graph store. Data and triples stay in the workspace.
An LLM drives the default path in four steps, and each step also has a manual workflow:
| Step | Action | What happens |
|---|---|---|
| 1 | Import Metadata | Fetches table and column metadata from Unity Catalog |
| 2 | Generate Ontology | The LLM designs entities, relationships, and attributes |
| 3 | Auto-Map | The LLM generates SQL mappings for every entity and relationship |
| 4 | Synchronize | Executes the mappings and populates the triple store |
OWL, R2RML, and SPARQL power the engine, and everything executes as Spark SQL, so users do not write RDF. Each domain can publish its graph to LLM agents over the Model Context Protocol (MCP), choosing which tools, datasets, and class actions it exposes, down to an ontology-only domain with a single describe tool. Domains are versioned and move through a draft, in-review, and published lifecycle with single-editor locking and an append-only review audit trail.
Features
- Ontology design: a drag-and-drop canvas (OntoViz), an AI assistant that edits through natural language, and imports of FIBO, CDISC, IOF, HL7 FHIR, and custom OWL/RDFS
- Pitfalls detector: 19 structural, logical, and semantic checks on the ontology
- Automatic mapping: the LLM generates SQL and column mappings that compile to R2RML, with schema-drift warnings
- Reasoning: OWL 2 RL inference, SWRL rules, and SHACL validation over the graph
- Exploration: two-phase search, N-hop neighbor expansion, cross-domain navigation, community detection, and cohort discovery
- GraphQL API: an auto-generated API over each graph
- MCP publishing: exposes domains to Cursor, Claude Desktop, or the Databricks Playground
- Pluggable graph backends: Lakebase (Postgres), Lakehouse Delta triple tables, Neo4j, or no backend for ontology-only domains
- Governance: versioned domains, review sign-offs with a configurable quorum, and OBX export/import between environments
