NVIDIA Agent Knowledge Foundation: How Auto Ontology Connects Agents to Business Data

NVIDIA announced Agent Knowledge Foundation on October 21 as part of its expanded Agent Toolkit. It brings together retrieval models, structured-data prediction and GPU libraries, with Auto Ontology providing a shared description of business data. For enterprise agents, that addresses a problem better search alone cannot solve.

An agent might find the right table and still answer the wrong question. Revenue can mean booked revenue, recognised revenue or cash collected. Returning valid SQL does not settle which definition the business intended.

What Agent Knowledge Foundation includes

The announcement groups Nemotron Retriever, Kumo structured-data models and CUDA-X libraries into a foundation for agents working with company information. Auto Ontology connects records to business concepts and relationships. OpenShell supplies the runtime controls around what an agent can access and do.

These are different jobs. Retrieval finds relevant evidence. An ontology describes entities, metrics and their relationships. Predictive models estimate outcomes. An optimisation library can turn predictions and constraints into a schedule or allocation.

NVIDIA says its supply-chain organisation is using the foundation through Palantir Foundry, Ontology and AIP. That is a concrete deployment example, not evidence that every integration is ready for every customer.

This is a collection of components, not a new all-purpose model. The announcement also does not mean that every underlying library was first released on October 21.

How Auto Ontology works

The Auto Ontology repository describes an open library that builds context from enterprise metadata, reconciles definitions across systems and exposes that context through MCP, a CLI and an API.

Its semantic catalogue sits in Postgres alongside the ontology. Natural-language questions can be translated into SQL against connected databases, while the system uses certified definitions and lineage to validate queries and provide provenance. The documented stack includes a web interface, a Python backend and an ingestion worker.

An ontology is a structured description of what data means. In an order workflow, that might connect a customer to an order, an order to a shipment and a shipment to a delivery status. The value comes from making those relationships explicit rather than asking each agent to infer them again.

The repository supports Docker Compose and Kubernetes deployment, with hosted or self-hosted NVIDIA NIM endpoints for inference. Auto Ontology is Apache-2.0 licensed. That does not make the database, inference or operational costs disappear.

What changes for an existing RAG system

NVIDIA's developer overview says the components can work with existing data platforms, ontologies, vector databases and agent frameworks. Teams can adopt them individually rather than replace their whole stack.

For a document-heavy assistant, retrieval may remain the main requirement. A workflow that combines policy documents with live orders and forecasts needs a more explicit separation between evidence, definitions and prediction.

Consider an agent asked which delayed shipments deserve attention. Finding the shipping policy is one step. Understanding which orders count as delayed is another. Predicting further delays and allocating limited delivery capacity require different tools again.

A shared definition layer can make those steps easier to inspect. It also creates something that needs an owner. If nobody maintains the meaning of a business metric, giving more agents access to it will not fix the ambiguity.

What still needs human review

NVIDIA describes Auto Ontology as producing a starting ontology that business experts test, refine and certify. It reports a 90% reduction in ontology creation time in its own testing. Treat that as a vendor result, not an expected saving for an arbitrary enterprise dataset.

The useful evaluation is whether the resulting definitions match the business, whether answers use the right records and whether the full workflow respects existing access controls. A convincing demonstration on clean data is only the beginning of that work.

For enterprise teams, the interesting shift is that agent quality increasingly depends on maintained business context outside the model. The next improvement may come from agreeing what a metric means before asking a more capable model to calculate it.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn complex enterprise AI products into clear stories that developers and decision-makers can understand and share.

Book a 30-minute call.