← Back to feed

NVIDIA has released Auto Ontology, an open-source Python framework that sits between AI agents and enterprise databases and checks what the agent writes against what the business has certified. The pitch in the repository is blunt: “a revenue metric answered faster is worth nothing if the number is wrong.”

What It Does

The library builds a graph-based semantic layer from enterprise metadata and exposes it to any agent harness through MCP, CLI and API. NVIDIA structures it around five verbs:

• Discover: generate ontologies from metadata, or connect existing ontologies and semantic models
• Ground: attach business meaning to metrics, entities and relationships, and reconcile conflicting definitions across systems
• Validate: check generated queries against certified definitions and lineage, and return provenance with every answer
• Expose: serve context over MCP, CLI and API
• Learn: feed certified corrections and usage back into the ontology

NVIDIA positions it as complementing semantic models customers already have, not replacing them.

The Stack

• Backend: FastAPI on Python 3.12+, with LangChain for natural-language-to-SQL
• Frontend: Next.js 16 with React 19, Better Auth for authentication and role-based access
• Storage: Postgres with pgvector holds the catalog, ontology and embeddings
• Data sources: Databricks, Postgres, Snowflake and DuckDB
• Deployment: Helm chart for Kubernetes or Docker Compose locally
• Default models via NVIDIA NIM: nvidia/nemotron-3-nano-30b-a3b for reasoning, nvidia/nemotron-3-embed-1b for embeddings, nvidia/llama-nemotron-rerank-vl-1b-v2 for reranking

The model endpoints are configurable per role, so other OpenAI-compatible or self-hosted NIM endpoints can be swapped in. Out of the box it needs an NVIDIA API key from build.nvidia.com.

What Is Missing

The repository points to a separate evaluation repo for running text-to-SQL benchmarks such as BIRD, but publishes no accuracy results of its own. Whether the validation layer cuts wrong-number rates on real enterprise schemas is untested in public. The licence is Apache 2.0, but the project is not accepting external contributions, so this is a code drop and not a community project yet.

Why It Matters

Text-to-SQL accuracy on academic benchmarks has stopped being the problem for enterprise buyers. The failure that costs money is a plausible query against the wrong table, or “revenue” computed with the wrong definition. A validation layer keyed to certified metrics targets that directly, and it fits NVIDIA’s habit of giving away software that increases demand for NIM endpoints and its GPUs.