WORLDTECHNVIDIA opens Auto Ontology, a semantic layer that grounds agent SQL in certified definitions
NVIDIA has released Auto Ontology as an open library under the Apache 2.0 license, a tool that derives a semantic layer from enterprise metadata so that agents can write SQL against certified business definitions. According to MarkTechPost, the layer is designed to sit between an agent and the data warehouse.
What the library provides
Auto Ontology builds its semantic layer from enterprise metadata. Agents draw on that layer when they compose SQL around certified business terms, and they receive the context through MCP, a command-line interface and an API.
The layer records where data resides and what each field signifies. It also indicates whether a field should be used, how it should be applied and where it came from.
Components and deployment options
The default language model is nvidia/nemotron-3-nano-30b-a3b, accompanied by Nemotron embedding and reranking models. The stack can run as a local Docker Compose setup or on Kubernetes via Helm. Inference is handled by hosted NVIDIA NIM endpoints or by self-hosted NIMs.
The MCP server documentation describes nine read-only tools. One returns the answer, the SQL and the rows as separate fields. No tool exists for browsing raw schemas, a deliberate choice that routes agent access to data through the semantic layer.
Five jobs and expert review
The library is organised around five jobs: discovering ontologies from metadata and connecting existing semantic models, grounding agents in defined metrics and relationships, validating generated queries against certified definitions and lineage, and feeding certified corrections back into the ontology. NVIDIA's developer page presents the output as a starting ontology that business experts must still test, refine and certify.
What remains unverified
The evaluation repository ships a test harness but no results, so performance claims cannot be checked from the release alone. The reporting outlet advises users to run the benchmarks themselves before trusting the system.
Open questions
- How accurate are the generated queries when measured with the evaluation harness?
- How much certified review by business experts is needed before the generated ontology is reliable?