Nvidia has released Auto Ontology, an open-source framework that gives AI agents business definitions to use when writing SQL against company data, Marktechpost reported on 1 November 2026. The Python project is available under the Apache 2.0 licence, allowing developers to build a shared context layer between agents and the databases they query.
Key points
- Auto Ontology builds a semantic layer from enterprise metadata and existing business definitions.
- Agents can request context through nine read-only MCP tools, as well as a command line and an API.
- Nvidia reports a 90% reduction in ontology creation time in its own testing; published text-to-SQL accuracy scores are not available.
- Developers can run the stack locally with Docker Compose or deploy it on Kubernetes with Helm.
Auto Ontology connects definitions across systems
An agent asked for a revenue number needs more than a list of tables: it needs to know which definition of revenue applies and which fields support it. Auto Ontology generates an initial map of entities, metrics and relationships from metadata, connects existing semantic models and reconciles definitions across systems. It then makes that context available to an agent before the agent writes a query, LavX reported on 3 November.
The map is an ontology, a description of how pieces of data relate to one another. Think of a warehouse with shelves labelled by different departments: knowing where a box sits helps, but finding the right box also requires agreement on what its label means. Auto Ontology uses metadata to assemble that map, while business experts test, refine and certify the definitions it produces. Its validation function checks generated queries against certified definitions and returns information about where an answer came from.
Checking a revenue figure would depend on whether its underlying definition had been certified. An agent could use that definition when writing the query and return the origin of its answer, rather than relying solely on a number that looks plausible.
The framework can also incorporate certified corrections and usage into the ontology. For companies connecting agents to existing data, that makes the business definitions a maintained part of the system rather than instructions attached separately to each question. Nvidia’s release follows its expansion of an Agent Toolkit for connecting agents to company data.
Nine MCP tools serve the context layer
Agents can reach Auto Ontology through an API, a command-line interface or MCP, the Model Context Protocol used to give an agent access to external tools. Its MCP server offers nine read-only tools. One returns an answer, the SQL behind it and the resulting rows as separate fields, with results capped at 100 rows. Another checks whether a compiled semantic layer and a live database connection are ready. The tools do not offer agents a way to browse raw database schemas, directing their requests through the semantic layer instead.
The stack has a Next.js 16 frontend for the web interface, a FastAPI backend for chat and natural-language-to-SQL requests, an ingestion worker that processes data and writes embeddings, and Postgres with pgvector for the catalogue and vector storage. Embeddings represent data in a form that lets the system retrieve related material; reranking then orders what has been retrieved. Through Nvidia NIM inference endpoints, the out-of-the-box setup assigns reasoning to nvidia/nemotron-3-nano-30b-a3b, embedding generation to nvidia/nemotron-3-embed-1b, and result ordering to nvidia/llama-nemotron-rerank-vl-1b-v2.
Developers can use Docker Compose for a local installation or Helm to deploy on Kubernetes. The default model setup requires a Nvidia API key, while model endpoints can be configured through environment variables. API tokens carry their owner’s role, with authorisation applied on each call. The project’s maintainers do not accept external contributions at this time.
Nvidia reports 90% faster ontology creation
Nvidia reports that ontology creation time fell by 90% in its own testing, according to Marktechpost. That measurement concerns building the context layer, a different task from judging whether an agent writes the right SQL or returns the right answer. Marktechpost reported that no text-to-SQL accuracy scores have been published.
A separate evaluation harness provides a way to test that second task. It runs a schema-only model alongside an ontology-grounded agent and scores whether their executed queries match the expected result, without using another language model as a judge. Its bundled test data includes BIRD Dev, with 1,534 questions across 11 databases, Mini-Dev with 500 questions, FDABench-Lite with 169 questions and a BEAVER configuration that defaults to a 100-question sample.
The harness documentation, as described by Marktechpost, says optional example questions and value hints go only to the ontology-grounded arm, so a comparison using them measures their contribution as well as the ontology’s. The tested configuration routes both arms through a Claude model on Nvidia’s inference gateway.
