Nvidia introduces auto ontology: an open framework for grounding AI agent sql queries

NVIDIA has launched Auto Ontology, an open-source semantic layer available under the Apache 2.0 license. This framework functions as a context layer situated between an AI agent and a data warehouse, enabling agents to generate SQL against certified business definitions derived from enterprise metadata. The system supports connections to various source databases, including Databricks, Snowflake, Postgres, and DuckDB. The architecture comprises four core services: a Next.js 16 frontend, a FastAPI backend utilizing LangChain for APIs, an ingestion worker, and a Postgres database with pgvector for catalog and vector storage. The platform offers 9 read-only tools via its MCP server, such as ask_question, which returns the answer, SQL, and rows, capped at 100 results. For evaluation, NVIDIA provides the auto-ontology-eval harness, which tests two arms—a schema-only LLM control and the ontology-grounded agent—using deterministic execution matching. Bundled datasets include BIRD Dev (1,534 questions across 11 databases) and Mini-Dev (500 questions). Local deployment can be managed via docker compose, while Kubernetes users receive a Helm chart. The system supports any OpenAI-compatible endpoint and requires an NVIDIA API key from build.nvidia.com for default models.