NVIDIA adds company-data tools and prediction models to its AI agent toolkit

NVIDIA is adding a new company-data foundation to its open-source Agent Toolkit, pairing tools for understanding business records with models for search and prediction. Announced October 21, the expansion combines that knowledge layer with controls over agent access and actions. NVIDIA says companies can adopt the foundation without moving data or rebuilding workflows.

From business records to predictions

The new Agent Knowledge Foundation starts with how an agent interprets company information. Its Auto-Ontology library translates raw data into recognizable business concepts, such as customers, products and orders, then maps their relationships. It sits alongside NVIDIA’s CUDA-X libraries, including cuDF, cuGraph, cuML, cuOpt and cuVS.

Search and prediction have different jobs within that foundation. Nemotron Retriever models help agents find relevant company information. The Kumo family works with structured business data—the records organized into tables and relationships—to forecast outcomes, score fraud risk and recommend actions, according to NVIDIA.

Kumo Structured includes models for tabular and relational prediction, time series and forecasting. Kumo Engine supplies an open library stack for developing and deploying structured-data foundation models. NVIDIA says developers can start with pretrained models, then fine-tune them on proprietary data when they need higher performance.

That is the distinction between the two model groups: retrieving information to answer a question versus using business records to predict an outcome. NVIDIA positions Kumo as a way to reduce the cost and complexity of building a separate machine-learning pipeline for every prediction task.

Understanding data does not set permission

The toolkit assigns the control role to OpenShell, its secure runtime—the environment in which an agent operates. NVIDIA describes it as the boundary that gives enterprises control over what agents may reach and do. The data foundation supplies business knowledge; OpenShell governs access and action.

NemoClaw blueprints provide reference designs for building secure, specialized agents. The newest blueprint focuses on agent optimization, offering a step-by-step starting point for improving an agent’s work over time. The expanded toolkit also includes tools and skills such as NeMo Relay and NeMo Compass.

Existing platforms provide the entry points

NVIDIA says enterprises can use Agent Knowledge Foundation while keeping their existing governance controls. Its deployment examples show several routes through software companies already use, rather than a single new interface:

• Palantir: NVIDIA’s supply-chain organization is using the foundation through Foundry, Ontology and AIP to deploy specialized agents, predict outcomes and support decisions.
• Databricks: Kumo Anomaly is paired with Databricks Genie Ontology to power agents that use business context.
• Snowflake: Customers are being enabled to use Kumo Engine and Kumo Relational open models inside Snowflake Notebook environments for predictive work.

For retrieval, NVIDIA cites gains in both quality and speed. It says Amazon OpenSearch Service improved relevance with fewer search steps and virtually no added latency using its embedding model. SAP’s fine-tuned reranker, which reorders search results, ran nearly three times faster. ServiceNow improved reranking quality while cutting latency by nearly a quarter. These are NVIDIA-supplied performance claims.

Kumo open models are available on Hugging Face and packaged for production inference through NVIDIA NIM on build.nvidia.com. Nemotron Retriever models are also available through Hugging Face, partner platforms and NIM. Those distribution options give developers access to the model components alongside the enterprise-platform integrations.