Box Plans Nemotron-Powered Retrieval for Enterprise AI Agents

Box plans to add Nvidia’s Nemotron embedding models to a search product for enterprise AI agents, aiming to help software find useful document passages faster and with lower processing costs.

The company demonstrated Box Content Retrieval at its developer conference Thursday, Nov. 5. The feature is designed to return passages relevant to an agent’s query or task. Box plans to roll it out to all Enterprise Advance customers in early 2027.

Box described the broader goal as making agent queries faster, more accurate and less expensive. Nvidia Vice President of Generative AI Software Kari Briski expects the feature to answer queries four times faster, with 20% fewer errors and 90% fewer tokens.

Those figures are Briski’s estimates. Box CEO Aaron Levie described the retrieval process as selecting excerpts for an agent instead of making it work through each full document. “We can send back chunks from the documents that are most relevant to the user’s query,” Levie said.

Nvidia’s Nemotron Embed project provides open-weight embedding models that prepare data for automated retrieval. The approach is designed to connect related concepts, which can help agents find relevant passages even when documents use different wording or languages.

Box identified legal work as one potential use, including helping lawyers find relevant clauses and precedents. The feature also uses Box’s existing metadata and permission controls. Those components are part of the planned system alongside the retrieval models.

Nvidia has made retrieval a focus of its open-model work. A Nemotron release in July topped the RTEB benchmark, which measures retrieval quality. Although the models are freely available, Box’s use of its own metadata and permission controls could make the complete system challenging to reproduce from scratch.

The Box plan adds to other enterprise applications of Nemotron, including Adobe’s planned integration of the models in Brand Concierge.