Box introduced Box Content Retrieval, a feature that brings Nvidia's Nemotron embedding system into its product to improve how AI agents search company content. Rather than requiring an agent to read every document, the system can return the passages most relevant to a query or task.
Nvidia generative AI software vice president Kari Briski told TechCrunch that the system is expected to answer queries four times faster, with 20% fewer errors and 90% fewer tokens. Box said the resulting searches should be faster, more accurate, and less expensive. The feature is scheduled to roll out early next year to all Enterprise Advance customers.
The product implements Nvidia's Nemotron Embed project, a collection of open-weight embedding models designed to improve automated retrieval. Embeddings map concepts rather than relying only on exact keywords, which can help an agent find related material even when a document uses different language. Box cited legal work as one example, where agents could locate relevant clauses and precedents across a document collection.
Although the Nemotron models are freely available, Box says its implementation also depends on the company's existing metadata and permission controls. CEO Aaron Levie described the retrieval feature as combining several technologies behind a single API. That integration is intended to preserve access rules while giving active agents more useful, focused context from enterprise files.