NVIDIA Expands Its Agent Toolkit to Ground AI in Enterprise Data

NVIDIA has expanded its open-source Agent Toolkit with the Agent Knowledge Foundation, a layer for grounding agents in company data while controlling their access and actions. The release combines ontology generation, retrieval, structured prediction and secure execution, addressing several systems that enterprise developers often have to assemble themselves.

Enterprise agents commonly use retrieval-augmented generation, or RAG, to search internal information before producing an answer. That approach works well for documents, but databases introduce schemas, relationships, permissions and forecasting tasks that vector search alone cannot handle. NVIDIA’s release post presents a broader stack for those workloads.

Five pieces from data to runtime

NVIDIA describes the toolkit as open source, though repositories and model releases may carry different licenses, dependencies and deployment requirements. Teams evaluating the stack will need to inspect those terms component by component.

Tables become business concepts

Auto-Ontology maps raw company data into concepts such as customers, products and orders, then records the relationships among them. An ontology gives an agent a business-level view of a schema, allowing it to connect an order row with the relevant customer, product and supplier without relying entirely on handcrafted prompts or application code.

The foundation uses several CUDA-X libraries: cuDF for dataframe processing, cuGraph for graph analytics, cuML for machine learning, cuOpt for decision optimization and cuVS for vector search. Together, these libraries cover the data transformations, relationship traversal and retrieval operations needed to build an agent-facing knowledge layer.

NVIDIA says the foundation can work with data in existing systems and preserve established governance controls. Developers should still trace the complete data flow because connectors may create indexes, embeddings, caches or other derived records that require their own retention and access policies.

Search and prediction share a layer

Nemotron Retriever handles embeddings and reranking for semantic search. Embeddings convert content into numerical representations, while a reranker reorders retrieved results so the most relevant records reach the agent first.

Kumo Relational can operate across joined tables, reducing the preparation required to flatten records into a single training table or feature store. That capability targets questions such as which customers may churn, which orders may arrive late or which products may run short, provided the underlying schema and training data support the prediction.

Runtime boundaries for tool-using agents

OpenShell supplies the secure runtime boundary for agents that query systems, invoke tools or change records. A runtime boundary can constrain access and actions; teams remain responsible for defining permissions, managing secrets, recording activity and testing how policies behave when tools fail.

NemoClaw blueprints provide reference designs for specialized agents, including a new blueprint for agent optimization. The templates give developers a starting architecture for evaluation and iterative improvement while preserving explicit controls around execution.

Early deployments and reported gains

NVIDIA names Dassault Systèmes, Palantir, Databricks and Snowflake among the early adopters and integration partners. The announced deployments include several concrete combinations:

• NVIDIA’s supply-chain organization uses the foundation with Palantir Foundry, Ontology and AIP to deploy specialized agents, generate intelligence from proprietary data, predict outcomes and support operational decisions.
• Databricks combines NVIDIA Kumo Anomaly with Databricks Genie Ontology for context-aware agents.
• Snowflake is enabling Kumo Engine and Kumo Relational models in Snowflake Notebook environments for predictive workloads.

Partner-reported performance

The announcement does not establish a shared benchmark across these deployments, which use different workloads, baselines and infrastructure. The results provide partner-specific evidence and require reproduction with each team’s data, latency targets and hardware configuration.

NVIDIA targets the agent stack’s missing middle

Agent frameworks such as LangChain and LlamaIndex coordinate models and tools, while data warehouses, vector databases and cloud platforms supply storage and managed services. NVIDIA is packaging ontology construction, GPU-accelerated processing, retrieval, structured prediction and secure execution into a stack that can sit between those systems and an agent application.

The approach has the clearest path for organizations already running NVIDIA GPUs and supported enterprise data platforms. CUDA acceleration can improve throughput, but it also introduces hardware dependencies that teams should measure against workload size, utilization and operating cost.

Checks before production

• Licensing: Confirm the license, redistribution terms and support status for every model and repository.
• Data flow: Identify where connectors, indexes, embeddings, caches and logs are stored.
• Schema quality: Test whether Auto-Ontology maps ambiguous names, keys and relationships correctly.
• Model quality: Evaluate retrieval and predictions with representative queries, tables and failure cases.
• Security: Verify tool permissions, network access, secret handling, audit records and isolation boundaries.
• Hardware economics: Benchmark latency, throughput and GPU utilization against simpler alternatives.

The stack is aimed at teams whose agents must search internal databases, traverse relational records, forecast business metrics or invoke tools under explicit controls. A focused proof of concept can test its central claims: whether Auto-Ontology captures the business schema, whether Nemotron and Kumo improve task quality, and whether OpenShell enforces policy under normal and failure conditions.