Box builds Nvidia's Nemotron embeddings into a new agent-first retrieval system

Box used its developer conference on November 5 to demonstrate Box Content Retrieval, a new feature that builds Nvidia's Nemotron embedding system into the core of its product so agents can answer queries using the embedded metadata Nemotron creates.

Nvidia vice president of generative AI software Kari Briski told TechCrunch that the system should answer queries four times faster, with 20% fewer errors and 90% fewer tokens consumed. Box frames the payoff as faster, more accurate and cheaper queries, three properties that map onto cost, experience and trust for enterprise search.

The rollout deserves precision: the feature will be available to all Box Enterprise Advance customers, but it ships early next year rather than on the day of the announcement.

Box CEO Aaron Levie described the mechanical shift. Instead of an agent reading through each document, the system can return the chunks most relevant to a user's query or the task the agent is working on, moving the heavy lifting of retrieval into indexing.

The foundation is Nvidia's Nemotron Embed project, which offers open-weight embedding models for automated retrieval. Embedding models have often been overlooked by frontier labs but are central to Nvidia's open-model effort: Nvidia's most recent Nemotron release in July ranked first on RTEB, a benchmark that measures retrieval quality.

Box's example is legal work. Lawyers can find relevant clauses and precedents more easily because Nemotron embeddings map underlying concepts rather than relying on keyword matches, so a search can surface related legal concepts even when the wording differs.

The moat is not the model. Nemotron's models are freely available, but the system leans heavily on Box's existing metadata and permission controls; Levie said the company is effectively packing five different technologies into one simple API, which makes it hard to rebuild from scratch.

For enterprise AI, the contest is shifting from how strong the model is to how well the retrieval layer works. Whether an agent gives usable answers inside a real document estate depends more on metadata quality, permission boundaries and embedding accuracy than on where a chat model sits in a leaderboard.

Two things to watch: the real query quality Box Enterprise Advance customers get once the feature ships early next year, and whether Nvidia can use integrations like this to make Nemotron embeddings a default foundation for enterprise retrieval.