Box previews Nvidia-powered search to help AI agents find relevant document passages

Box is preparing to give AI agents relevant passages from company documents rather than make them read through each file. At its November 5 developer conference, it previewed Box Content Retrieval, which integrates Nvidia’s Nemotron embedding technology. Box plans to roll the feature out to all Enterprise Advance customers in early 2027.

Finding concepts, then returning passages

The feature uses Nemotron to create embeddings: representations of document content that help a search system match underlying concepts, rather than only keywords. That gives the retrieval system a way to identify related material even when the wording—or the language—differs from the query.

Box CEO Aaron Levie described the next step as selecting the useful parts of those documents. Instead of asking an agent to read each document, the system returns the chunks most relevant to a query or task.

Box points to legal work as an example. A lawyer’s agent could use the system to locate relevant clauses and precedents, including material that expresses a related legal theme in different terms.

Nvidia’s projections for the system

Box’s promise is faster, more accurate and less expensive queries. Nvidia’s vice president of generative AI software, Kari Briski, gave TechCrunch specific expectations for speed, errors and token use. These are projections for the system, not established results from the customer rollout still ahead.

Nemotron’s retrieval technology has a separate benchmark record. A July release topped RTEB, a benchmark measuring retrieval quality. That result concerns the underlying model release; Briski’s figures concern expectations for the system being built into Box.

Open models, Box-specific infrastructure

The feature implements Nvidia’s Nemotron Embed project, which supplies open-weight embedding models for automated retrieval. Those models are freely available. But the planned Box service is not simply access to a model: it also relies heavily on metadata and permission controls already inside Box.

That existing infrastructure makes the complete system challenging to recreate from scratch, TechCrunch reported. Levie described Box as packing roughly five technologies into one simple API, the interface software uses to request the service. His description puts the integration—not just the availability of Nemotron’s model weights—at the center of what Box is offering.