NVIDIA Releases Auto Ontology: An Open Semantic Layer That Grounds AI Agent SQL

Every company trying to put an agent in front of its warehouse runs into the same wall. The model can write a query that runs fine and still returns the wrong revenue number, because it doesn't know which field the finance team actually trusts. Auto Ontology is NVIDIA's answer: an Apache 2.0 library that builds a semantic layer from enterprise metadata so agents query against certified business definitions instead of raw tables.

It matters because it's open, it speaks MCP, and it's built to sit under any agent harness. If this pattern sticks, the semantic layer becomes the control point for enterprise AI, and the question of who certifies the definitions becomes a real governance job.

We like the core design choice here. The MCP server exposes nine read-only tools and, on purpose, no tool for browsing raw schemas. Agents reach data only through the semantic layer, and every call carries the signed-in user's permissions. That's the right instinct. Most of the damage from data agents won't come from rogue models. It'll come from confident answers built on the wrong column, and forcing everything through certified definitions is a sensible guardrail.

We also appreciate that NVIDIA frames the generated ontology as a starting point that business experts still test, refine and certify. That's honest about where the human labor goes. Auto-generation can draft the map. It can't decide which definition of an active customer is the official one.

Now the eyes-open part. The headline number, ontology creation cut by 90%, is NVIDIA's own testing. There are no published text-to-SQL accuracy scores. NVIDIA ships an evaluation harness with BIRD and other benchmarks, but a harness is not a result. And the harness itself flags that few-shot exemplars and value anchors go only to the ontology arm, so any gain measured that way isn't grounding alone. Teams should read that as an invitation to run the numbers themselves before trusting it with a board deck.

There's also a quiet platform play. The defaults are Nemotron models on NVIDIA NIM endpoints. It accepts any OpenAI-compatible endpoint, which is good, but defaults shape behavior, and I'm watching whether the easiest path keeps routing enterprise data reasoning through NVIDIA's stack.

If this scales, the risk shifts rather than disappears. A wrong definition certified once gets reused by every agent that asks. Centralized meaning is powerful precisely because errors propagate cleanly. The correction loop that folds certified fixes back into the ontology will matter as much as the generation step.

Semantic layers aren't new. Tools like dbt MetricFlow and Cube define metrics in code, and Wren AI uses agent-assisted context files in Git. Auto Ontology's difference is generating the layer from metadata first and asking people to certify it, then serving it over MCP, CLI and API. It runs locally with Docker Compose or on Kubernetes via Helm, and connects to Databricks, Postgres, Snowflake and DuckDB.