NVIDIA Auto-Ontology: who decides what revenue means?
An open context layer can help agents use business definitions consistently. Someone still has to decide which definitions deserve to be trusted.
Ask two teams for last quarter’s revenue and you may get two defensible answers. One counts bookings. Another counts recognized revenue. An AI agent can produce perfectly valid SQL while choosing the wrong one.
That is why NVIDIA’s open-source Auto-Ontology interests me. The hard part of enterprise analytics often sits between the question and the query: agreeing on what the question means.
What it adds
Auto-Ontology builds a shared context layer from database metadata. It connects business terms to tables, relationships and SQL definitions, and provides tools for people to review and certify them. Agents can access that context through interfaces including MCP and REST. The code is Apache 2.0 licensed; running it still requires infrastructure and model inference. NVIDIA’s repository
Think of the ontology as a working dictionary that also knows how the data connects. Instead of reconstructing “recognized revenue” from column names each time, an agent can consult a definition the business has reviewed.
“What was last quarter’s revenue?”
- Compile proposed contextDatabase metadata → terms, relationships and SQL
- Review the business meaningA business owner resolves “bookings” vs. “recognized revenue.”
- Agent consults the shared definitionUse reviewed context; clarify ambiguous questions.
- Query and show the evidenceReturn the answer with SQL and steps people can inspect.
The ongoing work: keep definitions current.
Conceptual workflow, simplified from NVIDIA’s documentation. The revenue example is illustrative.
The part I would pay attention to
A consistent definition can still be consistently wrong. Generating a first draft of the ontology is useful. Deciding who can certify it, resolve conflicting definitions and update it when policy changes is the operating work.
I would start with one disputed metric, not a company-wide knowledge graph. Ask finance and operations to agree on its meaning, test questions that expose their differences, then see whether the agent handles those questions more reliably. Include questions it should clarify rather than answer.
NVIDIA’s companion evaluation compares a model given a schema with an ontology-enabled agent, scoring SQL by executing it against reference answers. That is a useful starting point. But the treatment also changes the agent’s tools and workflow, so any gain should not automatically be attributed to the ontology alone. Evaluation design
I would measure correct answers, appropriate clarification, review time and the cost of maintaining definitions. I would also check deployment defaults: NVIDIA documents that connection passwords are stored unencrypted in the catalog database when Vault is not configured. That matters before connecting real business systems. Connection documentation
My take: this is a promising way to make business meaning reusable across agents. Its value will depend on whether teams can keep that meaning trustworthy as the business changes.
Sources reviewed November 5, 2026. I have not run the framework or independently measured its accuracy. This is an analysis of the available project, not a claim about its first release date.
Follow the evidence
Sources & further reading
- NVIDIA Auto-Ontology: capabilities, model endpoints and Apache 2.0 license ↗
- NVIDIA: Auto-Ontology evaluation design ↗
- NVIDIA: source connections and credential storage ↗
Sources checked Nov 5, 2026. Research and drafting assisted by AI.
Suggest a correction