Box puts Nvidia's Nemotron at the core of a new agent-first retrieval layer

At its BoxWorks developer conference on November 5, Box introduced Box Content Retrieval, a feature that builds Nvidia's Nemotron embedding system into the core of the product. Instead of having an agent read through documents, the system answers queries using embedded metadata generated by Nemotron.

Nvidia supplied the numbers. Kari Briski, the company's VP of generative AI software, told TechCrunch she expects queries to be answered four times faster, with 20% fewer errors and 90% fewer tokens used. Those are vendor projections, and results will depend on how a given enterprise's content is structured.

Mechanically, Box chunks documents, precomputes embeddings and returns only the most relevant chunks when an agent asks a question. CEO Aaron Levie described the effect as dramatically boosting the relevance of the company's search technology, because agents no longer have to read every document to find the passage that matters.

The foundation is Nvidia's Nemotron Embed project, which publishes open-weight embedding models for automated retrieval. Major labs have largely overlooked the category, and Nvidia has made it a focus of its open model work; its July Nemotron release topped the RTEB benchmark, which measures retrieval quality.

Availability is scoped: Box Content Retrieval will be offered to all Enterprise Advance customers and rolls out early next year, which makes it an enterprise-tier capability rather than a default for every Box user.

The models are free to download, but Box argues the system is hard to recreate. It leans on Box's existing metadata and permission controls, and Levie said the company is effectively packing five technologies into one API. For regulated industries, inherited permissions may matter more than retrieval accuracy.

Legal work is the clearest example. Concept-level embeddings can surface related clauses and precedents across different wording or even different languages, the cases keyword search tends to miss. At the same event, Box also rolled out industry-specific solutions and controls over what AI agents can access and do inside Box.

The interesting part is distribution rather than the model itself. Open-weight embeddings are available to anyone, but bundling retrieval, permissions and an existing content graph into an interface enterprises can adopt is where Box sits. Watch the real latency and cost figures after the rollout begins next year, and whether agent-first retrieval extends beyond Box's own repository.