NVIDIA just open-sourced the answer to why your AI analyst keeps pulling the wrong revenue number.
The Summary
- NVIDIA released Auto Ontology, an open-source framework that builds semantic reasoning layers over enterprise data so agents know what metrics mean, where they live, and whether they can be trusted
- Addresses the core reliability problem in agentic analytics: speed means nothing if the answer is wrong
- Gives agents business context through metadata reconciliation and serves it via MCP, CLI, and API to any agent harness
The Signal
The enterprise AI problem isn't that agents can't generate SQL. It's that they generate the *wrong* SQL because they don't understand what "revenue" means in your three different databases. NVIDIA Auto Ontology attacks this at the metadata layer.
Auto Ontology builds knowledge graphs from enterprise metadata, reconciles conflicting definitions across systems, and exposes trusted business context to agent workflows. When your sales database calls it "gross revenue" and your finance system calls it "total bookings" and your data warehouse calls it "recognized revenue," the framework maps those relationships and tells the agent which one to use for which question.
"A revenue metric answered faster is worth nothing if the number is wrong."
The framework handles five jobs:
- Discover: Generate ontologies from existing metadata and connect semantic models
- Ground: Give agents certified business definitions for metrics and entities
- Validate: Check generated queries against trusted lineage before execution
- Expose: Serve context through Model Context Protocol and APIs
- Learn: Fold corrections back into the ontology so it improves with use
This is infrastructure for the agent economy. Not the sexy part where an agent books your travel. The unglamorous part where it needs to know that Q4 revenue in the CRM export doesn't match Q4 revenue in the earnings report because one includes deferrals and the other doesn't.
NVIDIA built this for natural-language querying of structured data, translating questions to SQL via NIM, their inference platform. But the real value is the semantic catalog layer. It sits between your agent and your databases, injecting business meaning into every query. When the agent asks "What was revenue last quarter?" the ontology layer knows which table, which column, which filters, and which caveats.
The benchmark angle matters: NVIDIA points to text-to-SQL benchmarks like BIRD for evaluation. That's the standard for measuring whether an agent can write correct queries. Auto Ontology is betting that query correctness isn't just about syntax. It's about semantics. You can generate perfect SQL that answers the wrong question if you don't understand the domain.
The Implication
This is open-source infrastructure for reliable agent analytics. If you're building agents that touch enterprise data, you need something like this or you're gambling every time the agent runs a query. The Model Context Protocol integration means it plugs into any MCP-compatible agent harness. NVIDIA isn't accepting external contributions yet, but they shipped the code under Apache 2.0. That's the signal: they want this to be standard plumbing.
Watch for two things. First, whether enterprises adopt this faster than they adopt the agents themselves. Second, whether the "fold corrections back into the ontology" learning loop actually works at scale. If it does, this becomes the shared context layer that makes agent-generated analytics trustworthy enough to act on.