The Short Answer
NVIDIA's Auto Ontology, released as open source under Apache 2.0 in November 2026, targets the confident wrong answer: an agent that runs valid SQL against the wrong meaning of a metric. It ships an evaluation harness instead of a headline accuracy score, because correctness depends on your data. ibl.ai applies the same principle to its ontology, where you own all the code and the data.
The release lives at github.com/NVIDIA/auto-ontology. The README frames the problem plainly: "a revenue metric answered faster is worth nothing if the number is wrong."
The viral version of this story called it "AutoAlign" and said it shipped "without a single benchmark." Both details are off, and the corrections make the point stronger, not weaker.
What did NVIDIA actually open-source?
NVIDIA open-sourced Auto Ontology, a Python library that builds a semantic layer over enterprise databases so AI agents answer from certified business definitions. MarkTechPost covered the release on November 1, 2026, and the Hacker News thread passed 290 points.
Per the repository, Auto Ontology connects to ten database engines, including PostgreSQL, Snowflake, Databricks, ClickHouse, Trino and DuckDB. It stores its ontology in Postgres with pgvector embeddings for retrieval.
The ontology holds three kinds of definition: terms (business concepts such as Customer or Invoice Line), reviewed SQL attributes (a margin, a rolling total), and relationships (how tables join). On Microsoft's Wide World Importers sample, one compilation produced 31 terms.
The workflow has three stages: the library generates a draft ontology, people certify it, and the agent answers from it. Agents reach it through a read-only MCP server with nine tools, a CLI, or an API.
Is NVIDIA AutoAlign a real product?
No NVIDIA product called AutoAlign was released. The open-source project is Auto Ontology. AutoAlign is a separate company whose hosted guardrails API, including a Factcheck rail, already plugs into NVIDIA's open-source NeMo Guardrails as a third-party integration.
The two are easy to blur because both sit near "catching wrong answers." They work at different layers, and only one of them can run entirely on your own infrastructure.
| The viral claim | What the primary source shows |
|---|---|
| "NVIDIA open-sourced AutoAlign" | NVIDIA open-sourced Auto Ontology (Apache 2.0). AutoAlign is a separate vendor with an API-key guardrail integration in NeMo Guardrails. |
| "The first tool built to catch the confident wrong answer" | NVIDIA's own NeMo Guardrails, open-sourced in 2023, already offers fact-checking and hallucination rails. Auto Ontology is new in where it acts: the meaning of the data, before the SQL runs. |
| "Shipped without a single benchmark" | No headline accuracy score is published, but a companion evaluation repository runs BIRD, BEAVER, FDABench-Lite and your own questions. |
Why do AI agents give confidently wrong answers on enterprise data?
AI agents give confidently wrong answers on enterprise data because valid SQL is not the same as a correct answer. A query can join the right tables, run cleanly and return a plausible number that uses a definition of "revenue" nobody in finance would sign.
Nothing in that chain throws an error. The model is fluent, the database is happy, and the result arrives with full confidence. That is the failure Auto Ontology is built around, and it is the same one we described in Worse Than Hallucination: Confidently Wrong.
Auto Ontology's answer has five steps per question: classify intent, retrieve candidate terms by embedding search, write SQL, check filter values and the query against the database, then return the answer with its SQL and the steps taken.
In the README's worked example, the agent reports revenue of $588.80 from 3 invoice lines and shows the full SQL and join path, so a person can check the number against the source rather than trust it.
How does Auto Ontology decide when not to answer?
Auto Ontology decides when not to answer with an answerability check. Before writing SQL, the agent asks whether the ontology covers every concept in the question, and reports the gap instead of guessing.
The repository's example asks for return on ad spend against a sales database that holds no advertising data. The check returns a coverage of 0.5, notes that no term covers "ad spend," and asks the user to clarify or connect another source.
That is the behavior that actually prevents a confident wrong answer. A system that says "this data cannot answer that" is worth more than one that is right 90% of the time and never tells you which 10% to distrust.
Why did NVIDIA skip a headline accuracy number?
NVIDIA skipped a headline accuracy number because the only number that matters is the one measured on your schema, your definitions and your questions. The auto-ontology-eval repository is built to produce that number, not to advertise one.
Each run asks the same model the same questions twice. The control arm gets the question plus a schema dump; the treatment arm gets Auto Ontology's retrieved context. One deterministic verifier scores both against gold SQL, with no LLM judge.
| Dataset in the eval harness | Questions | Engine |
|---|---|---|
| BIRD (Dev / Mini-Dev) | 1,534 / 500 | SQLite |
| BEAVER (dw sample) | 100 | MySQL |
| FDABench-Lite | 169 | SQLite |
| Wide World Importers | 41 | Postgres |
| Your own data | Your gold questions | Your database |
The harness also refuses to print a headline number when more than 5% of questions got no model output. A rate-limited run scores near zero and looks exactly like a weak model, so the guard stops a broken run from becoming a quoted statistic.
The last row is the one that counts. Public benchmarks tell you a method works somewhere; your own gold questions tell you whether it works on the definitions your finance, registrar or claims team actually uses.
Where does your data go when you run Auto Ontology?
Auto Ontology's services and catalog run on your own infrastructure: Docker Compose or a Helm chart, with the ontology in your Postgres.
Its four model roles default to NVIDIA-hosted NIM endpoints on build.nvidia.com, but accept self-hosted NIM or any OpenAI-compatible endpoint, so the whole stack can stay inside your perimeter.
One optional feature leaves the perimeter. Setting a KumoRFM key routes predictive questions to a hosted service that uploads connected tables, up to 20 by default. The README tells you to complete your organization's data review first, which is the right warning.
The README also notes that the backend and ingestion ports must stay on a private network, since the ingestion endpoints carry no authentication of their own. Teams in regulated sectors should read that section before any pilot.
Who should own the ontology your AI agents answer from?
The organization whose definitions it encodes should own the ontology your AI agents answer from. Certified terms, reviewed SQL attributes and the gold questions that prove them are institutional knowledge, built and reviewed by people.
That is the case for ibl.ai's Agentic Ontology: an MIT-licensed organizational ontology you self-host and expose to any authorized agent over MCP. Auto Ontology's Apache 2.0 license points the same direction, and the two ideas are compatible.
On ibl.ai you own all the code and the data, the platform stays model-agnostic across any LLM, and there is no per-seat pricing. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.
Teams that also want runtime guardrails on the agents reading that ontology can run them on NemoClaw, where ibl.ai deploys sandboxed agents with NVIDIA NeMo Guardrails inside your perimeter.
For the architecture behind a semantic layer you keep, see The Semantic Layer AI Agents Need — and Who Should Own It.
Frequently Asked Questions
Is NVIDIA Auto Ontology free to use commercially?
Yes. The repository is licensed under Apache 2.0, which permits commercial use. NVIDIA maintains the code and does not accept external pull requests, but issues and integration requests are open.
Does Auto Ontology replace a semantic model I already have?
No. The README says it complements the semantic models customers already own, and it imports and exports definitions in the Apache Ossie format.
Does it publish an accuracy score?
Not in the main README. The companion eval repository measures the accuracy difference with and without the ontology, on public benchmarks or on your own gold questions.