NVIDIA has released Auto Ontology, an open-source semantic layer framework designed to help AI agents write accurate SQL queries from enterprise metadata. Licensed under Apache 2.0, the library addresses the risk of incorrect data by establishing a trusted context layer that sits between agents and data warehouses. In NVIDIA testing, the framework cut ontology creation time by 90 percent.
The system relies on four core services, including a Next.js frontend, a FastAPI backend using LangChain, an ingestion worker, and a Postgres database utilizing pgvector. It supports source databases such as Databricks, Snowflake, Postgres, and DuckDB, while allowing integration with any OpenAI-compatible inference endpoints. Additionally, the framework exposes nine read-only tools through a Model Context Protocol server, enabling secure data access governed by user permissions.
While NVIDIA includes an evaluation harness with datasets like BIRD Dev and FDABench-Lite, the company has not published official text-to-SQL accuracy scores, leaving users to run their own benchmarks. The stack can be deployed locally via Docker Compose or on Kubernetes using Helm charts. Why it matters: This open-source framework provides enterprises with a standardized way to ground AI agent queries in certified business definitions, reducing data errors in automated analytics.