Box chooses Nvidia models to help AI agents search more accurately

Box chooses Nvidia models to help AI agents search more accurately

How Box Content Retrieval works

Instead of reading entire documents, an agent will receive passages that best match its query or task. Box CEO Aaron Levie said this should make search results more relevant to agents.

The system uses Nvidia’s Nemotron Embed project, which includes open embedding models for automated search. These models process data using large language models so agents can use it more effectively. TechCrunch reports that major labs often pay little attention to this class of models. But search has become one of Nvidia’s open-project priorities: in July, a new version of Nemotron took first place in the RTEB benchmark, which evaluates search quality.

Search by meaning and permissions

Box expects the system to be especially useful alongside active agents. The company gives lawyers as an example: they will be able to find relevant contract clauses and precedents more easily. Nemotron embeddings account not only for keywords but also for related concepts. Search can therefore find material on similar legal topics even when documents use different wording or language.

Nemotron models are available for free, but Box Content Retrieval also relies on Box’s existing metadata and permission settings. That makes the system difficult to reproduce from scratch. Levie said five technologies are combined in a single application programming interface.

Dataist’s take: Box combines Nvidia’s open models with its own metadata and access controls, so the result depends on more than embedding quality. The stated gains in speed, accuracy and token use are Nvidia’s expectations; the launch timing is known only in broad terms: early next year.