Any team that has pointed an AI agent at a data warehouse knows the failure mode. The agent returns a number. It is fast, fluent, formatted correctly — and wrong. A revenue figure that took four seconds instead of four hours is worth nothing if the definition of "revenue" came from the agent's best guess at a column name.
NVIDIA's answer, published on GitHub under Apache 2.0, is to stop letting the agent near the schema at all. Auto Ontology builds a semantic layer from enterprise metadata — the entities, metrics, relationships and lineage the business already has opinions about — and exposes only that layer to the agent. The repository states the problem without decoration: reliability gates agentic analytics.
The library does five jobs. It discovers ontologies from metadata and connects existing semantic models. It grounds them, reconciling definitions that disagree across systems. It validates generated queries against certified definitions and returns provenance with every answer. It exposes the result over MCP, a CLI and a REST API. And it learns, folding certified human corrections back into the shared context.
The deliberate omission is the interesting part. The MCP server ships nine read-only tools, and among them there is no way to browse a raw schema. That is a design decision, not an oversight: agents reach the data through the semantic layer or not at all. `ask_question` returns the answer, the SQL and the rows as three separate fields rather than a paragraph of prose, and caps results at 100 rows. Behind it the published OpenAPI spec lists 87 operations, with every call carrying the signed-in user's permissions.
The stack is unremarkable in the best way. A Next.js 16 frontend handles UI and auth on port 3000; a FastAPI backend with LangChain runs chat and natural-language-to-SQL on 3001; an ingestion worker writes embeddings on 3002; Postgres with pgvector stores the catalog, the ontology and the vectors on 5432. Credentials can go to HashiCorp Vault — worth noting, because without it connection strings are stored in plaintext on the catalog row, password included. Source databases can be Databricks, Postgres, Snowflake or DuckDB. Inference defaults to NVIDIA's own Nemotron family through NIM endpoints, but any OpenAI-compatible endpoint works.
Then there is the evidence gap. NVIDIA reports that ontology creation time fell 90% in its own testing. It publishes no text-to-SQL accuracy scores at all. A separate harness, auto-ontology-eval, ships in a sibling repository — BIRD Dev with 1,534 questions across 11 databases, a 500-question Mini-Dev set, FDABench-Lite, a 100-question BEAVER sample — and scoring is deterministic execution match rather than an LLM judge, which is the right call. But the harness is an invitation to produce results, not the results themselves.
That harness also documents its own two caveats, which is rarer and more credible than a clean number would have been. Optional few-shot exemplars and value anchors are fed to the ontology arm only, so any measured delta includes those inputs and not grounding alone. And its tested configuration routes both arms through a Claude model on NVIDIA's inference gateway — a detail that matters to anyone planning to reproduce the comparison on their own models.
The project also is not accepting external contributions; pull requests from outside the maintainer team will not be reviewed. For an Apache 2.0 release positioned as infrastructure, that is a notable posture: open to read, open to run, closed to steer.
What NVIDIA has actually shipped is a well-built argument about where the reliability problem lives. The five functions map cleanly onto the real failure modes of agentic analytics, and refusing to expose raw schema is the sort of constraint that only comes from having been burned. The question the repository leaves open is the only one that decides adoption: not whether the layer is well-architected, but whether it makes the numbers right. That answer now has to come from the teams running the harness themselves.
