NVIDIA Auto Ontology: Open semantic layer for AI agents

What changed

NVIDIA has released Auto Ontology, an open library under the Apache 2.0 license that sits between AI agents and data warehouses. The framework builds a semantic layer from enterprise metadata, allowing agents to write SQL against certified business definitions instead of raw database schemas. The core problem it addresses is straightforward: a fast revenue answer is worthless if the number is wrong.

The stack consists of four core services. A Next.js 16 frontend handles the web UI and authentication on port 3000. A FastAPI backend with LangChain runs chat, natural-language-to-SQL translation, and catalog APIs on port 3001. An ingestion worker writes embeddings on port 3002. Postgres with pgvector stores the catalog, ontology, and vectors. Source databases can be Databricks, Postgres, Snowflake, or DuckDB. The default LLM is nvidia/nemotron-3-nano-30b-a3b with Nemotron embedding and reranking models, though any OpenAI-compatible endpoint works.

The framework enables five core functions. Discovery generates ontologies from metadata. Grounding gives agents trusted meaning for metrics and relationships. Validation checks queries against certified definitions. Exposure serves context through Model Context Protocol, CLI, and API. Learning folds certified corrections back into the ontology.

Why it matters

Enterprise AI agents face a critical reliability problem. A database query that executes without errors can still return wrong answers. An agent might join tables incorrectly, select the wrong metric, or misinterpret what a field means. These errors often produce plausible-looking numbers that pass casual inspection. Auto Ontology addresses this by placing certified business definitions between the agent and the raw data.

This matters for anyone deploying agents to query databases at scale. Hallucinations are not just embarrassing; they are costly in finance, healthcare, and operations where incorrect metrics drive real decisions. By grounding agents in a semantic layer built from enterprise metadata, the tool reduces the surface area for these failures.

The framework also handles definition reconciliation across systems, a problem every large organization faces. Different teams often have different names and definitions for the same concept. Auto Ontology discovers these conflicts and allows business experts to certify a single trusted definition.

Deployment flexibility matters as well. Teams can run Auto Ontology locally via Docker Compose or scale it on Kubernetes with a Helm chart. NVIDIA publishes the backend as an NGC container, and an NVIDIA API key covers the default models.

What to test

Before adopting Auto Ontology, run your own benchmarks. NVIDIA ships a separate evaluation repository called auto-ontology-eval that runs two test arms: a schema-only LLM baseline and the ontology-grounded agent. Scoring uses deterministic execution match with no LLM judge. Bundled datasets include BIRD Dev (1,534 questions across 11 databases), Mini-Dev (500 questions), FDABench-Lite (169 questions), and BEAVER (100-question default sample).

However, NVIDIA has not published text-to-SQL accuracy scores. The evaluation repository ships a harness and test datasets, but not results. This is a significant gap. You must run the benchmarks yourself against your own ontologies and data to trust the tool in production.

Test whether the 90% ontology creation time improvement NVIDIA reports in self-reported testing holds on your schema. Measure how long it takes to generate, refine, and certify ontologies for your databases. Also verify that the MCP tools work correctly with your agent harness. The library exposes 9 read-only tools, including ask_question (returns answer, SQL, and rows separately, capped at 100 rows) and check_readiness (confirms a compiled semantic layer and live database connection exist).

Verify whether the semantic layer actually reduces hallucinations on your use cases. Compare agent accuracy querying raw schemas versus querying the ontology. Also audit the permission model. Authentication runs through OAuth and every call carries the signed-in user’s permissions, but verify this prevents unauthorized data access in your environment.

Finally, test the ingestion pipeline on your source databases. Confirm embeddings are written correctly and that value anchors and few-shot exemplars flow to the ontology arm as documented.

The conclusion

Auto Ontology is a credible open-source effort to solve a real problem: grounding AI agents in certified data definitions. The architecture is sound, the deployment options are flexible, and the semantic layer concept is proven. However, the lack of published accuracy scores on standard benchmarks is a significant gap. NVIDIA has not shown public evidence that the tool materially improves SQL correctness on datasets like BIRD.

This is not a flaw in design, but a gap in communication. You need to benchmark it yourself on your data before production use. The evaluation harness is available; the results are not. If you run Auto Ontology internally and see consistent accuracy gains, it is worth adopting. If you skip the benchmarking step and deploy it assuming the 90% ontology creation improvement translates to fewer hallucinations, you will regret it.

Watch for published benchmarks from NVIDIA or the community. Also track whether other agent tools integrate Auto Ontology’s MCP interface, which would expand its reach. The semantic layer pattern is spreading, and Auto Ontology is now a solid open reference implementation.

Related reading

• agent-desktop: Open-source agent computer use
• StreamVibe: Google’s video agent for timely help
• dsh-superagent: Router-worker coding agent for parallel GitHub tasks