Box integrates Nvidia Nemotron for agent content retrieval

Box brings Nemotron embeddings into enterprise search

Box has announced Box Content Retrieval, a feature that integrates Nvidia’s Nemotron embedding system into the core of its product. The capability is designed to help agents answer queries using embedded metadata, rather than working through every document in full.

The company plans to make the feature available to all Enterprise Advanced customers in early next year. Nvidia vice president of generative AI software Kari Briski said the approach is expected to deliver answers four times faster, reduce errors by 20%, and use 90% fewer tokens.

From document reading to targeted retrieval

Box CEO Aaron Levie said the system changes what an agent receives in response to a request. Instead of requiring the agent to read each document, Box can return the sections that are most relevant to the query or task.

Box Content Retrieval implements Nvidia’s Nemotron Embed project, which provides open-weight embedding models for automated retrieval. Embeddings preprocess information for use by agents, supporting retrieval based on relationships between concepts rather than only matching the exact words entered in a query.

That distinction is particularly relevant to legal work, an example used by Box. Lawyers could find relevant clauses and precedents even when a search uses language different from that in the underlying material, because the embeddings are intended to map related legal concepts.

Open models meet Box’s enterprise controls

Nvidia said its most recent Nemotron release, published in July, ranked first on the RTEB benchmark for retrieval quality. The project’s models are freely available, but Box’s implementation also depends on its existing metadata and permission controls.

Levie described the product as combining roughly five technologies in a single API. That combination means the retrieval layer is not simply an embedding model: it is tied to the content information and access rules maintained within Box.

The deployment also arrives as Nvidia’s role in AI infrastructure remains central to the market dynamics described by Nvidia’s AI infrastructure market pressures while enterprise software vendors seek practical ways to make agents more effective over controlled business data.

What enterprise teams should assess

For organizations using document-based agents, the announced figures make retrieval a measurable operational question: teams can evaluate answer speed, error rates and token consumption alongside the relevance of returned content. They should also validate that access permissions and metadata behave correctly in the workflows where agents will operate.

Box’s planned Enterprise Advanced rollout gives customers a route to test whether concept-based retrieval can reduce the amount of material sent to an agent without losing the context needed for a task. The practical business implication is that enterprises can assess retrieval quality and governance together before committing broader agent workflows to production.