NVIDIA Auto Ontology Puts a Human-Certified Semantic Layer Between Agents and Enterprise Data
Agents that query enterprise data now have a way to fail loud instead of wrong: certified definitions and explicit refusal on low coverage turn silent data corruption into a visible, fixable event. That's the difference between a demo and a tool you can attach to a production reporting pipeline.
Enterprise data agents routinely produce runnable SQL with wrong numbers because column names like `revenue` don't carry tax treatment or metric ownership. NVIDIA's Auto Ontology tackles that by generating a context layer from metadata: it extracts business terms, audited SQL expressions, and table relationships, then requires a human to test and certify them before any agent uses them. The three-step Generate-Certify-Answer loop writes everything to PostgreSQL with pgvector embeddings.
The key behavior is honest refusal. When retrieval candidates can't support a query, the agent reports a coverage score and says it can't answer, rather than fabricating a result. The system exposes read-only MCP tools, a REST endpoint, and Skills; it supports ten databases including PostgreSQL, Snowflake, ClickHouse, and DuckDB, and works with any OpenAI-compatible model endpoint.
Deployment caveats matter: ingestion endpoints on ports 3001/3002 have no authentication, all default passwords must be changed, and the project does not accept external pull requests. Also, despite the repo description mentioning deterministic query generation, SQL is still written by an LLM; only the definitions are human-certified and validated.
The honest refusal mechanism treats low evidence as a first-class output, not an error state. That's a more practical reliability design for enterprise query agents than forcing a confident answer, because it converts silent wrong results into a checkable coverage number.
The repo description's 'deterministic query generation' claim is softer than it sounds: SQL is still LLM-generated, only the definitions are human-certified and the output is validated. Teams that require fully reproducible queries need to verify this distinction themselves.
Auto Ontology is deliberately additive, designed to coexist with existing semantic models rather than replace them. That positions it as a compatibility layer for agent consumption, which makes adoption easier but also means its value depends on how much certification effort a team is willing to invest.