Box taps Nvidia’s Nemotron for new agent-first content retrieval system

Box used its developer conference Thursday to unveil Box Content Retrieval, a feature that puts Nvidia's Nemotron embedding system at the center of how the platform serves agents. Rather than having an agent page through entire documents, Box returns the chunks most relevant to the query or task at hand, backed by metadata Nemotron creates during indexing.

Nvidia's VP of generative AI software, Kari Briski, told TechCrunch she expects queries to come back four times faster, with 20% fewer errors and a 90% reduction in tokens used. The feature will be available to all Box Enterprise Advance customers and is scheduled to roll out early next year.

Box CEO Aaron Levie framed the change as a relevance upgrade for agent-facing search: "Instead of an agent having to kind of read through each document, we can send back chunks from the documents that are most relevant to the user's query or whatever the task is that the agent is doing." The underlying technology is Nvidia's Nemotron Embed project, which publishes open-weight embedding models designed to improve automated retrieval — an area frontier labs have largely passed over even as Nvidia has pushed it through its open-source model work. Nvidia's most recent Nemotron release, in July, ranked #1 on the RTEB benchmark for retrieval quality.

Box's pitch centers on domain use cases where wording diverges from intent. The company pointed to lawyers locating relevant clauses and precedents, arguing that Nemotron embeddings map underlying concepts rather than keywords, so a search can surface related legal ideas expressed in different language. Levie also stressed that the offering is not simply the open models repackaged: Nemotron itself is freely available, but Box layers in its pre-existing metadata and permission controls, and Levie said "we're probably packing in five different technologies into one simple API."

The questions ahead are mostly about delivery and defensibility. Whether the stated speed, accuracy and token savings hold across large, permission-heavy enterprise repositories will become clearer as the early-next-year rollout reaches Enterprise Advance accounts. Equally open is how much of the advantage comes from Nemotron versus Box's metadata and permissions stack — and whether rivals embedding the same open-weight models can close the gap without that accumulated content governance. The value also appears tied to how actively agents operate over Nemotron-processed data, so adoption may track agent maturity inside customer organizations as much as the retrieval benchmarks themselves.