Nvidia Open-Sources Auto-Ontology to Boost AI Agent Accuracy

AI agents can retrieve and query data, but correctness matters more than speed in enterprise analytics—a wrong answer delivered quickly destroys value. Nvidia’s Auto-Ontology creates a semantic layer that sits between agents and databases, mapping business terminology to columns, storing validated SQL for consistent metrics, and capturing table relationships. This gives agents reliable context about what data exists, what it means, and how to safely combine it.

The system automates ontology generation from database metadata, then enables human review and certification through a web UI. Developers can validate whether questions are answerable before execution, ensuring agents don’t fabricate results for questions the data can’t address. The ontology gets exposed to agents through standard protocols like MCP and REST APIs, with generated SQL and reasoning steps included for transparency and audit trails.

As enterprises scale AI-driven analytics, a certified semantic layer could significantly reduce both answer errors and operational overhead. Nvidia benchmarked Auto-Ontology against datasets like BIRD and FDABench-Lite, measuring accuracy differences between the semantic layer approach and agents working from raw database schemas.