Nvidia announced on Wednesday during the GTC conference in Berlin the launch of a new package called "Agent Knowledge Foundation" as part of the Agent Toolkit, aimed at providing enterprise artificial intelligence agents with missing data context and improving their decision-making capabilities.
Components of the New Package and Retrieval Technologies
The new package includes a collection of open-source components comprising "Nemotron Retriever" models for retrieval, the new "Auto-Ontology" library, the "Kumo" family of models dedicated to structured data, alongside "CUDA-X" data processing libraries. The company explained that this system allows data to remain in its current storage locations while complying with pre-existing access rules and permissions enforced within companies.
Adele ElHallak, Vice President of Agentic AI Products at Nvidia, noted that large language models face extreme difficulty in searching through complex tables and files without clear structural context. According to the company's tests, the Nemotron Retriever ranking and rerrieval model increased search result relevance by 65% when tested on an open-source OpenSearch environment, thereby reducing the number of required search steps and maintaining minimal latency.
Transition from Traditional RAG to Ontology and Table Processing
The package relies on the Auto-Ontology library to build knowledge graphs from corporate tables and connect raw data to business concepts such as customers, products, and orders, while providing a REST API and a Model Context Protocol (MCP) server for agent queries. This development capitalizes on the expertise of the startup Illumex, which Nvidia acquired in February for approximately 75 million dollars.
The package also integrates Kumo models for relational data, which Nvidia acquired in June. These models aim to eliminate traditional pipelines dedicated to each predictive task such as classification and sentiment analysis, enabling automated prediction across structured tables.
Strategic Partnerships with Data Platforms
Nvidia emphasizes that it does not seek to compete with major data platforms such as Snowflake, Databricks, or Palantir, but rather relies on the integration of its infrastructure with their platforms to manage Access Control Lists (ACLs) and security. Nvidia's supply teams use these technologies via the Palantir Foundry platform, while Databricks connects the Kumo anomaly detection model with the Genie Ontology context layer, and Snowflake allows running Kumo models inside its notebooks.
Why This News Matters
This launch represents a practical step to overcome the current limitations of enterprise artificial intelligence agents, which often suffer from an inability to understand the interconnected context of distributed corporate data. By integrating open-source models with ontology layers, Nvidia provides developers with flexible tools to improve agent accuracy without forcing organizations to transfer their sensitive data or completely restructure it.