← Back to blog
AiAbout 6 min read

NVIDIA Wants Agents to Know What a Number Means Before They Query It

Published Nov 14, 2026
NVIDIA Wants Agents to Know What a Number Means Before They Query It

The pitch for text-to-SQL has always been speed. Ask a question in English, get a number in seconds, skip the analyst queue. The pitch has a hole in it that everyone who has actually shipped one of these systems knows about: a fast wrong number is worse than a slow right one, because nobody double-checks the fast one.

NVIDIA's Auto-Ontology, which picked up roughly 4,100 stars in its first week on GitHub and now sits above 12,500, starts from that hole. Its README puts the argument in one line: agents can retrieve documents and query databases, but reliability gates agentic analytics, and a revenue metric answered faster is worth nothing if the number is wrong.

What it builds is a context layer that sits between the agent and the data. Not a vector index, not a bigger prompt. An ontology: what data exists, what it means, whether it should be used, how it should be used, where it came from, and what it resembles.

Five verbs

The library is organised around five capabilities, and the naming is unusually honest about what each one is for.

Discover generates enterprise ontologies from metadata, connects ontologies and semantic models you already have, and builds the knowledge graph. This is the part that has to cope with reality, which is that most companies have three definitions of "active customer" and no record of which one the dashboard uses.

Ground gives agents trusted business meaning for metrics, entities and relationships, and reconciles inconsistent definitions across systems. In practice this is the hardest of the five and the least glamorous: somebody has to decide that the finance definition wins, and the tool's job is to make that decision visible and enforceable.

Validate checks generated queries against certified definitions and lineage, and returns provenance with every answer. That provenance line is the actual product. A number with a definition and a lineage attached can be argued with. A number without one can only be trusted or ignored.

Expose serves the context to any agent harness through MCP, a CLI, or an API. Learn folds certified corrections and usage back into the ontology so the shared context improves with each question.

That last one is the only genuinely novel piece of the loop. Most semantic layers are documents that rot. This one treats the correction a human makes as training data for the layer itself.

What the stack looks like

For a project with this much ambition, the deployment story is refreshingly ordinary, which matters if you are the person who has to run it.

The frontend is Next.js 16 with React 19, TypeScript and Tailwind 4, plus Better Auth and Prisma. The semantic catalogue and the ontology live in Postgres alongside the catalogue they describe. Natural-language querying runs through NVIDIA NIM. There is a web UI for chat, analysis, the data catalogue, analytics and settings.

Locally, the quick start is Docker Compose against Wide World Importers, Microsoft's sample sales database. You need Docker with Compose v2, uv, and an NVIDIA API key from build.nvidia.com. The sequence is: clone, create .env, bring up Postgres, seed the sample database from the separate evaluation repo, then docker compose up -d --build. Open localhost:3000, sign in with the admin credentials in .env, go to Settings → Semantic Compilation and turn it on. Ingestion runs at startup; compilation begins immediately and reports "Semantic compilation is ready" when it finishes.

On Kubernetes there is a Helm chart. Licence is Apache 2.0, and the README states plainly that the project is not accepting external contributions, which is worth knowing before you plan a fork-and-patch strategy.

The uncomfortable part of the design

An ontology is a claim about what is true in your company. Building one automatically from metadata is fast, and the output is a draft of that claim, not the claim itself.

The README says it complements semantic models customers already own rather than replacing them, and it says it reconciles definitions. Reconciliation is a political act. If marketing and finance disagree about what counts as a new customer, no amount of graph construction settles it. The tool can surface the conflict and record the decision. It cannot make the decision.

That is where the work actually is, and it is why the library ships a reconciliation step instead of a magic wand. Agentic analytics keeps disappointing for a boring reason: models can write SQL, and nobody had written down which SQL is correct.

Why this reaches beyond dashboards

The same problem shows up in creative operations, which is where this gets relevant to anyone running image and video pipelines.

Ask an agent "what did we spend on generation last month" and watch what happens. Does that include retries? Does it count drafts that were discarded? Is a failed generation billed? Is the number per asset or per request? Three teams will answer those four questions differently, and each will build a dashboard that is internally consistent and externally wrong.

An ontology does not fix the ambiguity. It makes it explicit, which is the precondition for fixing it. The alternative, which is what most companies currently do, is let each dashboard carry its own silent set of definitions and hope nobody compares them.

How to tell whether you need it

The honest test is not "do we have messy data", because every company has messy data. The test is whether anyone has already written down the correct answer to a question two teams currently answer differently.

Pick the metric your company argues about most. Ask three people in three departments to define it. If the definitions match, you do not have an ontology problem, you have a documentation problem, and a wiki will do. If they do not match, no amount of model capability will rescue you, because the agent's query will be correct against one definition and wrong against two others, and it will not know which.

The second test is whether you have a certification process. Auto-Ontology's value comes from the Validate step, and Validate needs something to validate against. If nobody in the company is authorised to say "this is the finance-approved definition of gross margin", the library will faithfully serve you a graph of three competing truths.

The honest summary

Auto-Ontology is an infrastructure project with an unglamorous thesis: agents are not limited by reasoning, they are limited by not knowing what the words in the question refer to inside this particular company.

Twelve thousand stars in a week is a signal about demand, not adoption. The people starring it are the ones who have already been burned by a confident wrong number. They are the right audience. Whether they will do the reconciliation work that makes the tool useful is a separate question, and it is the one that will decide whether this becomes infrastructure or another repository in a bookmarks folder.

Related articles