NVIDIA Auto Ontology: Validating Text-to-SQL with DB Metadata-Based Ontologies

NVIDIA has released Auto Ontology, a Python-based open-source framework that automatically generates ontologies from enterprise DB metadata and provides a semantic layer to validate AI agents' Text-to-SQL.
It validates SQL using business term definitions and lineage rather than schema-name-based guessing, and includes provenance in its responses; NVIDIA claims that its own tests reduced ontology creation time by 90%.
The full-stack application structure is built with Next.js 16, FastAPI, and Postgres + pgvector, supporting Databricks, Postgres, Snowflake, and DuckDB, and uses NVIDIA NIM endpoints by default.
It provides nine read-only tools via an MCP server, allowing queries based on SSO login without DB credentials, while intentionally excluding raw table exploration tools to prevent bypassing the semantic layer.
It uses a LangGraph-based Text-to-SQL graph as backend logic and provides optional inspection options such as interactive clarification and DB probes for SQL literals and join paths.
It does not accept external contributions, and standalone use of the Python library is not possible, making full-stack deployment mandatory; in air-gapped environments, preparation of a self-hosted NIM or OpenAI-compatible endpoint is required.
It applies the Apache-2.0 license, and direct verification is required through the separate evaluation tool Ontology SQL Eval; unlike SQRL, it modifies only the context and validation procedures without changing the model.

Key point