NVIDIA Agent Knowledge Foundation

NVIDIA Agent Knowledge Foundation is a collection of open models and accelerated libraries for connecting enterprise data to AI agents. It combines document retrieval, structured-data processing, predictive models, graph analysis and decision optimization. NVIDIA presents the components as independently adoptable tools that work with existing data platforms and agent frameworks.^[1]

Components

The collection covers several stages of an agent's work. Each component has its own interfaces and deployment requirements.^[1]

The tools address different problems. cuDF supplies data-processing operations, cuVS provides vector-search primitives, cuGraph operates on graph structures, and cuOpt takes a mathematical objective and constraints. Their separate documentation describes these interfaces.^[2][3][4][5]

Retrieving evidence and preparing data

Nemotron Retriever supplies the parsing, embedding and reranking stages of retrieval-augmented generation. Parsing extracts document content; embedding represents content for search; reranking orders retrieved candidates. The collection also includes structured-data tools, so its scope extends beyond document search.^[1]

cuDF accelerates operations such as loading, joins, aggregation, sorting and shuffling. It uses Apache Arrow's columnar memory format and offers Python and C++ interfaces, alongside integrations that let supported operations run through an existing data engine. The cuDF documentation describes CPU fallback for operations or data types that an integrating engine cannot run on the GPU.^[2]

cuVS provides vector-index construction and search. Its documented applications include semantic search, recommendation and RAG. It exposes APIs for several programming languages and integrates with existing search libraries and databases. A vector-search library supplies indexing and search operations; deploying an application still requires its surrounding data storage and service logic.^[3]

Business meaning and predictive analysis

Auto Ontology provides a starting description of enterprise concepts and relationships, which business experts can inspect and certify. This adds a semantic interpretation to the data an agent uses. NVIDIA includes Kumo for pretrained predictions over structured data, and cuML for building custom machine-learning pipelines.^[1]

cuGraph provides graph analytics through high-level Python APIs, a NetworkX backend and lower-level native libraries. Its documentation also separates traditional graph analytics from graph-neural-network integrations. These interfaces let applications work with connected entities and their relationships.^[4]

From predictions to constrained decisions

cuOpt is an open-source GPU-accelerated decision-optimization engine. Its documentation covers linear programming, vehicle routing and quadratic programming, with some other solver capabilities designated beta. It also supplies agent skills for formulating, solving, debugging and explaining optimization problems.^[5]

A prediction and an optimization result answer different questions. A model can estimate an outcome from data; an optimizer searches for a decision under a specified objective and constraints. NVIDIA's collection includes both kinds of tool, alongside the retrieval and processing components that supply their inputs.^[1][5]

Integration and evaluation

NVIDIA's example applications include supply-chain analysis, portfolio optimization and machine-learning pipelines. They combine subsets of the collection rather than requiring every component in each application.^[1]

Performance and answer quality must be evaluated separately. GPU acceleration changes the execution of supported operations; it does not establish that an agent selected the right business definition, retrieved complete evidence or formulated the right optimization problem. The foundation's component documentation defines different operations and evaluation settings, so a result for one component should retain its workload and configuration when reported.^[1][2][3][5]

Continue reading

• AAI AgentsAn AI agent is a software system that selects and performs actions in an environment in pursuit of an objective.Linked in the introduction
• NNVIDIA Auto OntologyNVIDIA Auto Ontology, styled Auto-Ontology in its repository, is an open-source service that supplies business definitions and database context to AI agents.Mentioned twice above
• NNVIDIA Agent ToolkitNVIDIA Agent Toolkit is NVIDIA's umbrella name for its collection of open models, agent blueprints, libraries, skills and runtimes for building and running AI agentsAlso in Enterprise AI
• NNVIDIA AI EnterpriseNVIDIA AI Enterprise is an end-to-end, cloud-native software suite sold by Nvidia as a paid subscription for developing and deploying production artificial intelligence and data analytics.Also in Enterprise AI
• NNVIDIA RivaNVIDIA Riva is a GPU-accelerated software development kit and family of containerized inference services for speech and translation AI, built by NVIDIA.Also in Enterprise AI
• HHarveyHarvey is an artificial intelligence platform built for the legal industry, developed by Counsel AI Corporation and headquartered in San Francisco.Also in Enterprise AI

More to read

• Sierra AI
• Glean (company)
• Hebbia
• Decagon (AI company)
• Agentforce
• OpenClaw Enterprise

Topics:AI AgentsEnterprise AINVIDIA

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7NVIDIA. Building an Agent Knowledge Foundation. Component overview and example workflows, accessed November 4, 2026.
2. ^1 ^2 ^3NVIDIA. cuDF: GPU-Accelerated Structured Data Processing. Architecture, interfaces and integration behavior, accessed November 4, 2026.
3. ^1 ^2 ^3NVIDIA. cuVS: GPU-Accelerated Vector Search and Data Clustering. Indexing, search and integrations, accessed November 4, 2026.
4. ^1 ^2NVIDIA. cuGraph documentation. Library structure and supported interfaces, accessed November 4, 2026.
5. ^1 ^2 ^3 ^4NVIDIA. cuOpt. Solver scope and agent integration, accessed November 4, 2026.