NVIDIA has open-sourced Auto Ontology, an Apache 2.0-licensed semantic layer framework designed to standardize how AI agents query data warehouses. The framework sits between autonomous agents and enterprise databases like Snowflake, Databricks, Postgres, and DuckDB, providing standardized metadata to ensure generated SQL adheres to certified business logic.
The open-source repository includes a Next.js frontend, a FastAPI backend using LangChain, an ingestion worker for vector embeddings, and a Postgres/pgvector catalog. The platform deliberately restricts agents from directly browsing raw database schemas, requiring all interactions to run through a Model Context Protocol (MCP) server with role-based user permissions.
To benchmark correctness, NVIDIA released an accompanying evaluation harness called auto-ontology-eval using NeMo Gym. The framework evaluates query generation accuracy against standard datasets like BIRD Dev and FDABench-Lite using deterministic execution matching rather than subjective LLM judges.
Why it matters
Enterprise engineers get an open architecture to stop AI agents from hallucinating incorrect SQL on production databases.
Adopting semantic layers like Auto Ontology reduces schema exposure and enforces deterministic enterprise permissions for agentic tools.
Standardizes Model Context Protocol (MCP) tool integration for connecting LLMs securely to backend data warehouses.
Source: marktechpost.com



