NVIDIA has released Auto Ontology, an Apache 2.0-licensed semantic layer framework that constructs a context layer from enterprise metadata to ground AI agents in certified business definitions for SQL generation. The system comprises four core services: a Next.js 16 frontend (port 3000), a FastAPI backend with LangChain for chat, NL-to-SQL, and catalog APIs (port 3001), an ingestion worker for embeddings (port 3002), and PostgreSQL with pgvector for storing the catalog, ontology, and vectors. It supports Databricks, PostgreSQL, Snowflake, and DuckDB as source databases. Models are configured via environment "triplets" (endpoint, key, model) with distinct roles for reasoning, non-reasoning, embedding, and reranking; the default LLM is Nemotron-3-Nano-30B-A3B served via NVIDIA NIM, though any OpenAI-compatible endpoint is accepted. The MCP server exposes nine read-only tools—including ask_question (returning answer, SQL, and rows capped at 100) and check_readiness —with OAuth authentication and user-scoped permissions; the OpenAPI spec publishes 87 operations. NVIDIA reports a self-claimed 90% reduction in ontology creation time but publishes zero text-to-SQL accuracy scores. The separate auto-ontology-eval harness runs deterministic execution-match benchmarks in NeMo Gym using BIRD Dev (1,534 questions), Mini-Dev (500), FDABench-Lite (169), and BEAVER (100-sample), comparing a schema-only LLM control against the ontology-grounded agent; however, optional few-shot exemplars and value anchors apply only to the ontology arm, and both arms route through a Claude model on NVIDIA's inference gateway. Deployment options include Docker Compose, a Kubernetes Helm chart, and an NGC container for the backend.