databrickslabs/ontobricks

What it solves

Enterprises have huge amounts of data in Databricks tables, but the meaning behind that data — relationships like "a Customer holds Contracts" or business rules — is usually locked in people's heads. OntoBricks turns those tables into a knowledge graph that captures meaning, so it can be explored, reasoned over, and exposed to AI agents. It removes the need for separate graph databases, RDF expertise, or long integration projects.

How it works

OntoBricks runs as a Databricks App and uses the services you already have: Unity Catalog for metadata, SQL Warehouse for execution, and AI Gateway or Model Serving for LLM calls. You go through four main steps: 1) Import metadata from Unity Catalog, 2) Generate an ontology — the LLM suggests entities, relationships, and attributes from your metadata, 3) Auto-Map — the LLM writes SQL mappings for each entity and relationship, and 4) Synchronize — run the mappings to populate the triple store. Under the hood it uses standards like OWL, R2RML, and SPARQL, but you don't need to know them. You can also design ontologies manually using a visual canvas, and there's a "No LLM" mode for strict opt-out. The graph can be built on Lakebase (Postgres), Lakehouse (Delta tables), Neo4j, or no backend at all. Every domain has a DRAFT/IN-REVIEW/PUBLISHED lifecycle with locking and audit trails. You can publish your ontology to LLM agents via the Model Context Protocol (MCP), and query the graph via GraphQL or SPARQL.

Who it's for

Data engineers, data architects, and analytics teams using Databricks who want to bring semantic meaning to their data. It's also for anyone who wants to expose governed, explainable business logic to AI agents or LLM applications. You don't need deep ontology or RDF skills because the AI handles most of the heavy lifting, but you can go manual if you prefer full control.

Highlights

• Turns Databricks tables into a knowledge graph in four clicks, powered by LLM automation.
• Uses open standards (OWL, R2RML, SHACL, SPARQL) but hides the complexity.
• Multiple graph backends: Lakebase (default), Lakehouse, Neo4j, or ontology-only with no backend.
• Publish any domain to AI agents over MCP (e.g., Cursor, Claude Desktop).
• Governance features: versioning, lifecycle statuses, single-editor locking, append-only audit trail.
• Includes ontology pitfalls detection, SHACL data-quality checks, SWRL business rules, and cohort discovery.
• Runs entirely inside your Databricks workspace — data and triples never leave.