Nvidia's Auto Ontology Claims a 90% Cut in Build Time and Publishes No Test, While Taking No Outside Contributions

Nvidia's Auto Ontology, an Apache-2.0 Python framework that builds a semantic layer over company databases for AI agents, says it "reduced ontology creation time by 90% in NVIDIA testing." The Nvidia developer page carrying that line gives no baseline, test set or method, and the GitHub repository it links to has no performance numbers either.

The repo stood at 2.4k stars and 181 forks when fetched on Nov. 3. The README says: "This project is currently not accepting external contributions."

What it does

An ontology, here, is a map of what a company's tables and columns mean and how they join, so an agent asks about "revenue" and gets the right table instead of a guess. The README calls Auto Ontology "a Python framework for building semantic, graph-based reasoning layers over data," built from enterprise metadata and served to agents through MCP (the Model Context Protocol agents use to call tools), a command line and an API.

Nvidia's page says it "generates a starting ontology from enterprise data and documentation" that business experts then test, refine and certify, and that it "complements existing ontology platforms." The 90 percent claim is therefore about the first draft, not the certified result.

What running it requires

The local stack needs Docker with Compose v2, or Kubernetes through Helm. The backend needs Python 3.12 or later and Postgres with pgvector. Source databases can be Databricks, Postgres, Snowflake or DuckDB.

The default model path is Nvidia's hosted API catalog, which takes an nvapi- key. Self-hosted NIM endpoints (Nvidia's packaged inference containers) are supported. The README does not state a GPU requirement. Nvidia's NGC container page lists the backend image at version 1.1.0, updated Oct. 13, 2026, under the Apache License 2.0.

The issue tracker is a to-do list

The repo shows 22 open issues and 11 open pull requests. Every one of the 22 issues was opened by the same account, cafzal, between Sept. 13 and Oct. 8. Read as a list of titles, which is all this report read, it describes the unfinished work.

Issue	Opened	Title
#259	Sept. 13	Preserve certified ontology edits across reset and rebuild
#261	Sept. 13	Qualify Q&A latency, cost and quality
#267	Sept. 14	Model import ID and natural-name conflicts return HTTP 500
#271	Sept. 14	Databricks optimized extraction loses nullability
#334	Oct. 4	Honor explicit requested forecast origin on the native prediction path

Issue #259 matters most, since certification is the step that is supposed to make the output trustworthy, and its title implies certified edits can be lost on a reset. Issue #261 says Q&A latency, cost and quality still need to be qualified, which sits beside a headline number about build time. The README's benchmark section does not carry results either: it points to a separate BIRD text-to-SQL evaluation repository.

What is and is not established

The repo shows 1,354 commits on main, so this is active work, and Apache-2.0 permits commercial use. What it does not establish is whether the 90 percent figure holds outside Nvidia's own test, or how much expert time certification then takes. Outside developers can file issues but, by the README's terms, will not see them reviewed or merged.

For how agent tooling behaves once installed, see Agent-Desktop 1.0's macOS bug that types into the frontmost app and the cost of running AI agents across JavaScript frameworks. The number to watch is the Q&A qualification in issue #261.