NVIDIA Auto Ontology is an Apache-2.0 framework that builds semantic context over enterprise metadata, validates generated queries and exposes that context to agents through MCP, CLI and APIs.
NVIDIA has released Auto Ontology as an open-source framework for building semantic, graph-based reasoning layers over enterprise data. The project targets a problem that is becoming central to agentic analytics: an agent can generate SQL quickly, but speed is irrelevant if it uses the wrong metric, table, join or business definition.
Auto Ontology is designed to sit between enterprise metadata and the agents that need to reason over it. It discovers or connects ontologies, reconciles business definitions, validates generated queries against trusted semantics and lineage, and exposes that context through MCP, CLI and API interfaces.
What is NVIDIA Auto Ontology?
Auto Ontology is an Apache 2.0 open-source project that builds a semantic and graph-based context layer over structured enterprise data.
According to NVIDIA’s repository, the framework is intended to help agents answer five questions before acting on data: where the data lives, what it means, whether it should be used, how it should be used and where it came from.
The project currently includes natural-language querying of structured data, an ontology/catalog layer stored in Postgres, authentication and RBAC, a web interface, and deployment paths through Docker Compose or Helm.
Why this matters for Text-to-SQL agents
A Text-to-SQL model can produce syntactically valid SQL while still being wrong in business terms. It may use a non-authoritative table, reconstruct a KPI incorrectly, join at the wrong grain or ignore lineage.
Auto Ontology addresses the context problem rather than treating SQL generation as an isolated model task.
Discover: build or connect enterprise ontologies from metadata.
Ground: give agents trusted definitions for metrics, entities and relationships.
Validate: check generated queries against certified definitions and lineage.
Expose: make the semantic layer available through MCP, CLI and APIs.
Learn: feed certified corrections and usage back into the shared context.
Auto Ontology also exposes an MCP server
The project includes an MCP server so compatible agent harnesses can query enterprise data while inheriting the authentication and permission checks of the main application layer.
Its documented ask_question capability runs the project’s full Text-to-SQL flow and can return the answer, generated SQL and result rows separately. That is useful because an agent can inspect and reuse the SQL without parsing it out of prose.
The architecture is broader than an LLM prompt
The repository currently uses a Next.js frontend, FastAPI services, Postgres with pgvector, and optional Vault integration. NVIDIA NIM endpoints are used for reasoning, embeddings and reranking by default, although endpoint and model settings are configurable.
This makes Auto Ontology closer to a semantic infrastructure component than a prompt-engineering recipe.
What Auto Ontology does not prove yet
The repository points users to a separate evaluation project for benchmarks such as BIRD. The main project page does not provide a universal Text-to-SQL accuracy improvement that can be applied to every enterprise dataset.
That distinction matters. Ontologies and semantic layers can reduce ambiguity and provide stronger validation, but reliability still depends on the source metadata, business definitions, model, schema, permissions and evaluation dataset.
Why this direction is strategically important
Agentic analytics is increasingly moving away from the idea that a larger model can infer every business rule from raw schemas.
The architecture is becoming layered:
That is closely aligned with the production architecture we describe in our Text-to-SQL guide. The model remains important, but business semantics, retrieval, validation and permissions determine whether the answer is actually usable.
How this relates to thaink²
thaink² approaches data agents with the same separation of responsibilities: the Data Analyst Agent reasons over enterprise data, while semantic context, permissions, tools, evaluation and AgentOps surround the model.
Auto Ontology is not a thaink² component and this article is not a product comparison. It is notable because a major infrastructure vendor is open-sourcing a semantic layer specifically for the class of problem enterprise data agents must solve.
For teams evaluating this architecture, the questions to test are concrete: does the ontology reflect authoritative business definitions, does it improve schema selection and metric consistency, can permissions be enforced end to end, and can the system prove where an answer came from?
What to watch next
independent Text-to-SQL evaluation on real enterprise schemas;
how well existing semantic models can be connected rather than regenerated;
governance of corrections and ontology changes over time;
performance and operational cost of the multi-call Text-to-SQL flow;
adoption of the MCP interface by external agent runtimes.
Frequently asked questions
What is NVIDIA Auto Ontology?
An open-source framework that builds semantic and graph-based context layers over enterprise data for grounding and validating agent queries.
Is Auto Ontology open source?
Yes. NVIDIA publishes it under Apache License 2.0.
Does Auto Ontology generate SQL?
The current project includes a natural-language-to-SQL flow and returns generated SQL and results.
Does Auto Ontology support MCP?
Yes. The project includes an MCP server for compatible agent harnesses.
Does it guarantee accurate Text-to-SQL?
No. Reliability still depends on metadata, business definitions, models, permissions and evaluation data.
