Nvidia bets open models can give enterprise AI agents the context they’re missing

Nvidia's latest announcement at the GTC Berlin conference introduces its Agent Knowledge Foundation to its Agent Toolkit, providing enterprises with open-source components to help AI agents better understand their company data. This system maintains data in its original location, adhering to existing access rules, without competing with existing enterprise data platforms.

Nvidia's vice president of product for agentic AI, Adel El Hallak, explains that while large language models (LLMs) are good at predicting next tokens, they struggle when presented with a single spreadsheet containing multiple tabs, as they have difficulty finding specific information. To address this, Nvidia's new system utilizes its Nemotron Retriever models, an Auto-Ontology library, structured data models, and CUDA-X data processing libraries.

These components work together to enhance retrieval accuracy and reduce the number of lookups required, ultimately decreasing latency. The Auto-Ontology library, built from a company's own tables, maps raw data to business concepts like customers, products, and orders. Nvidia's acquisition of Illumex, an Israeli startup specializing in automated ontology creation, for $75 million in February, further strengthens its position in this area.

Kumo, a startup Nvidia acquired in June, is known for building foundation models for relational data, with several of its models released as open weights on Hugging Face. Nvidia's approach is to partner with enterprise data platforms like Snowflake and Databricks, rather than compete with them. El Hallak emphasizes that the company aims to provide seamless integration with existing systems and security features, allowing enterprises to maintain control over their data and access controls.