NVIDIA Auto Ontology: an open semantic layer for AI agents

Felix Laurent 3 Nov 2026 4 min read

AI agents

NVIDIA has open-sourced Auto Ontology, a library that tells AI agents what company data actually means. It builds a semantic layer from database metadata, so an agent writes SQL against approved business definitions instead of guessing from table names. MarkTechPost covered the release on November 1, 2026, and the code is on GitHub under the Apache 2.0 license.

"A revenue metric answered faster is worth nothing if the number is wrong," the project's README says. That is the problem most AI analytics tools still have.

The problem Auto Ontology tries to fix

Ask an AI agent for last quarter's revenue, and it will answer in seconds. But a company database can hold several columns that look like revenue: gross, net, booked, recognized. If the agent picks the wrong one, the chart still looks clean, and nobody notices.

The issue is not SQL. Modern models write valid queries well. The hard part is knowing which table and field match a business term, and whether the data is fit to use.

Auto Ontology calls its answer a context layer. It sits between the agent and the data warehouse. It tells the agent where data lives, what it means, whether and how to use it, and where it came from.

How the semantic layer works

The README describes five steps:

Discover: generate an ontology, a map of business terms, from database metadata, or connect existing semantic models;
Ground: give agents trusted meanings for metrics and entities, and reconcile conflicting definitions across systems;
Validate: check generated queries against certified definitions and lineage, and return provenance with every answer;
Expose: serve the layer to any agent through MCP, a command-line tool or an API;
Learn: fold certified corrections back into the ontology, so each answered question improves it.

The key word is certified. Auto Ontology drafts the ontology automatically, but business experts test, refine and approve it, NVIDIA says. A user who asks a question gets the answer, the SQL that produced it and the rows behind it.

The way agents reach this layer is where the design gets strict.

Why agents cannot see raw tables

Auto Ontology ships an MCP server, so tools like Cursor and Claude Desktop can query company data through it. The server exposes nine tools, and all of them only read. "Nothing modifies the glossary, the catalog, or the underlying databases," the documentation says.

The main tool, ask_question, takes a plain-language question and returns an answer with SQL and rows. Others search the business glossary, explain a term or show which columns a term maps to.

What is missing is deliberate. There is no tool to browse databases, schemas or raw columns. In the docs' words, a catalog browser "invites an agent to reason about raw tables instead." The agent has to go through business terms.

The MCP server holds no database credentials itself. It calls the Auto Ontology API and acts with the signed-in user's permissions.

What you need to run it

The stack is built from familiar parts. The frontend uses Next.js 16 and React 19. The backend runs on FastAPI and LangChain, and Postgres with pgvector stores the catalog, the ontology and the embeddings.

It connects to four source databases: Databricks, Postgres, Snowflake and DuckDB. It runs locally with Docker Compose or on Kubernetes with a Helm chart.

By default, inference goes to NVIDIA NIM endpoints with Nemotron models: Nemotron 3 Nano 30B for reasoning, plus Nemotron embedding and reranking models. Any OpenAI-compatible endpoint can replace them through environment variables.

Auto Ontology is part of NVIDIA's Agent Knowledge Foundation. That set of open models and libraries also includes Nemotron Retriever, cuDF and cuVS for preparing and searching enterprise data.

What the release does not prove yet

NVIDIA's one performance claim is that Auto Ontology reduced ontology creation time by 90% in its own testing. No baseline or method is published. MarkTechPost summed up the release as "strong engineering, thin public evidence."

There are no accuracy scores either. The README points to a separate evaluation repository for text-to-SQL benchmarks such as BIRD, but it ships a test harness, not results.

Two practical limits are worth knowing. Database connections added through the interface are stored unencrypted in Postgres unless HashiCorp Vault is set up, the README warns. The project also accepts no outside contributions: issues and pull requests from outside NVIDIA's team will not be reviewed.

Developers are interested anyway. The repository had about 2,270 GitHub stars and 170 forks on November 3.

Does Auto Ontology solve AI analytics

Not by itself, but it targets the right problem. An agent that writes perfect SQL against the wrong column is still wrong, and Auto Ontology forces every query through approved business terms. Its MCP server makes that the only path: nine read-only tools and no way to browse raw tables.

The open question is how well it works. NVIDIA has published a 90% time saving with no baseline and no accuracy scores at all. Until teams run the BIRD harness on their own data and share results, Auto Ontology is a promising design, not a proven fix.