NVIDIA's Kumo models let agents predict from your tables. The ontology they depend on still needs an owner
NVIDIA's Agent Knowledge Foundation adds an Auto-Ontology library and Kumo models for relational, tabular and time-series prediction. Agents move past document search; meaning becomes the weak point.
Listen to this article · 5 min
AI-generated narration of the full article.
Ask a typical enterprise agent which of this quarter’s customers are likely to stop buying, and it will do what it was built to do: search. It finds documents that mention churn, a dashboard export, perhaps last year’s board presentation, and writes a fluent paragraph. What it cannot do is look at the relationships between customers, orders, products and support tickets and estimate, customer by customer, what happens next.
That gap is what NVIDIA addressed on October 21 with the Agent Knowledge Foundation, an expansion of its open source Agent Toolkit. The pieces:
- Auto-Ontology, a new library that translates raw company data into business concepts, such as customers, products and orders, and maps how they relate.
- Kumo, a new family of open models for structured data: Kumo Tabular, Kumo Relational, Kumo Time Series and Kumo Forecast, plus an anomaly model that Databricks pairs with its Genie Ontology.
- Nemotron Retriever models for search, and CUDA-X libraries such as cuGraph for graphs and cuOpt for optimization.
NVIDIA’s own supply chain organization uses the stack through Palantir’s Foundry and AIP. Snowflake customers can run the Kumo Engine and Kumo Relational inside Snowflake notebooks on GPU compute pools, so data, training and predictions stay in the account; Snowflake names Affirm, NOV, VideoAmp and Yieldmo as users. Several other vendors say they are building on it.
Four questions an agent may need to answer
The useful way to read the announcement is as a list of different questions. Retrieval answers only the first.
| Component in NVIDIA's stack | Who must validate it | |
|---|---|---|
| What do our documents say? | Retrieval: Nemotron Retriever, rerankers | Content owners; retrieval evals |
| What do our terms mean, and how do they relate? | Auto-Ontology | The business owner of each definition |
| What is likely to happen to this customer, order or machine? | Kumo predictive models | Data science, with backtests on your history |
| Which action is best under these constraints? | cuOpt optimization | Operations, against real constraints |
The second and third rows are what is new. A relational model that learns from tables, keys and the timing of events can estimate churn, fraud, demand or the next purchase without someone first flattening the data into one wide table, which is how much predictive modeling in enterprises still works. Snowflake’s example predicts which active customers will buy again in the next 90 days from four tables and a few lines of Python.
A prediction is a tool to call, not a thought to have
The architectural lesson is that an agent should not estimate the probability of churn by reasoning about it in text. It should call a model trained to estimate it, receive a number with known accuracy, and reason about what to do with that number. We made the same point when Google put a forecasting model inside BigQuery: not every enterprise AI problem needs a language model doing the thinking.
That division also makes the system testable. A Kumo model can be backtested on last year’s data, its error measured and its drift monitored, like any predictive model your risk team already reviews. A language model’s improvised estimate cannot.
The ontology is the new weak point
Auto-Ontology is the most consequential and the most fragile piece. Mapping raw tables to business concepts automatically is attractive because building ontologies by hand is slow. It is risky for the same reason. A generated ontology will decide that a “customer” is a row in one table, that “revenue” is a particular column, that two entities are related because a key says so. Those are business decisions, and in most companies they are contested.
We have written that business definitions are becoming shared agent infrastructure, and that in regulated finance the ontology may matter more than the chatbot. An automatically generated ontology does not remove that work; it produces a first draft of it. Treat the output as a proposal that the owner of each definition approves, versions and can change, not as configuration the platform team accepts by default.
Cautions before adopting
- Vendor results are vendor results. NVIDIA’s examples of faster reranking and better relevance come from partners and are not independently measured.
- Prediction is not causation. A model that predicts churn well does not explain why customers leave, and an agent acting on a prediction can change the behavior it predicted.
- Check what is open and how it is licensed before planning on it: the Agent Knowledge Foundation combines new components with ones NVIDIA released earlier, and availability differs by piece.
- Keep the decision with the system that owns it. An agent that predicts a customer will churn and offers a discount is making a commercial commitment. That should pass through the same approval as any other discount.
The direction is right. Enterprise agents that can only search documents will keep producing confident paragraphs about questions that need numbers. The work that comes with it is also clear: someone has to own what the words in the ontology mean.