NVIDIA Auto Ontology

NVIDIA Auto Ontology is an open-source Python framework that builds a semantic, graph-based layer over enterprise data. It generates an ontology from database metadata, lets people certify it, and exposes it to agents so that their data queries use the same business definitions. Its README describes it as a way to make analytics reliable: agents answer with the SQL and steps behind each result, so the answer can be checked. It sits alongside the semantic models a customer already owns and uses Apache Ossie (a separate Apache project) to keep the ontology portable.

Repository snapshot (GitHub API, 2026-11-22)

ItemDetail
RepoNVIDIA/auto-ontology, Python, Apache-2.0
Created2026-04-27
Activitylast push 2026-11-22; no formal release is published (the latest-release endpoint returns 404)
Size14,322 stars, 1,272 forks

What it does (README)

  • Discover: generates an enterprise ontology from metadata, connects existing ontologies and semantic models, and builds knowledge graphs for AI use.
  • Ground: gives agents business meaning for metrics, entities and relationships, and reconciles inconsistent definitions across systems.
  • Validate: tests a draft with real questions before experts certify it, checks generated queries against the catalogue, and returns the SQL and steps with each answer.
  • Expose: serves context to any agent harness through MCP, a CLI and an API.
  • Learn: folds certified corrections and usage back into the ontology.

How the ontology is built

The ontology answers six questions, each with its own element:

QuestionElementIn Auto Ontology
What does it mean?ConceptA term, such as Customer, with a description and synonyms
Where is it?MappingColumn attributes that tie each term to tables and columns
What’s connected?RelationshipJoins between tables, including bridge tables: declared keys plus inferred links checked against the data
How do I calculate it?MetricReviewed SQL attributes reused under a term
Where does it come from?LineageEach term and attribute links to its source tables and columns
Should I use it?CertificationExperts certify names, descriptions and attributes

Generation runs in stages: the ingestion service reads tables, columns, keys, views and query history, embeds tables and columns in pgvector, and optionally tags PII. A model then proposes foreign-key candidates and junction tables, writes column descriptions and names the terms. People refine the draft, and semantic compilation is switched on in the web app’s settings.

Quickstart (README)

The README runs the stack with Docker Compose against the Wide World Importers sample database, Microsoft’s sample sales database. It needs Docker with Compose v2, uv and an NVIDIA API key from build.nvidia.com. The setup has you set several passwords and API keys in .env before starting the containers. In the example, a question about revenue for one product in one state in one month returns the value, the SQL and the resolved terms. A question the data cannot answer, such as return on ad spend in a dataset with no advertising data, gets a clarification request instead of a guess.

Where it fits

  • Text-to-SQL and analytics: the closest fit. The point is consistent metrics and checkable queries, not free-form generation.
  • Agents: the MCP server lets an agent check whether a question can be answered before it writes SQL. See model-context-protocol.
  • Compared with other ontology tooling: a LinkedIn post from November 2026 sets it against Neo4j’s Ontology Graph Specification and AWS’s Context Ontology Accelerator, and describes NVIDIA’s approach as a semantic layer over enterprise tables and SQL metrics, without OWL inference. That comparison is one author’s reading and is not checked here. For the wider landscape see knowledge-graph-and-ontology-platforms and ontology.

Source limits

  • The README’s leaderboard badges (#1 on BEAVER, #2 on BIRD-Interact) are the project’s own statements. Neither leaderboard’s results were checked here.
  • The README does not state licence terms beyond Apache-2.0, nor the cost of the NVIDIA API calls the stack makes.
  • No formal release is published, so the version is the repository’s current state, not a tagged version.

Sources