Box builds Nvidia's Nemotron embeddings into agent-first content retrieval

At its developer conference on Thursday, Box announced Box Content Retrieval, a feature that builds Nvidia's Nemotron embedding system into the core of its product. According to TechCrunch, the new capability is aimed at AI agents: rather than having an agent read through entire documents, the system returns the chunks most relevant to a query or task, using metadata generated by Nemotron.

Nvidia has put concrete numbers on the expected gains. 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. The token reduction matters as much as the speed: for enterprises running agents across large document sets, inference cost scales with token volume, so a 90% cut implies substantially cheaper operation.

The feature will be available to all Enterprise Advance customers and is set to roll out early next year.

An implementation of Nemotron Embed

The feature is an implementation of Nvidia's Nemotron Embed project, which publishes open-weight embedding models designed to improve automated retrieval systems. Embedding models use large language models to pre-process data so that agents can work with it more effectively — a category that, as TechCrunch notes, frontier labs have largely overlooked, but which Nvidia has made a focus of its open-source model work. Nvidia's most recent Nemotron release in July ranked first on the RTEB benchmark, which measures retrieval quality.

Box CEO Aaron Levie framed the change as a boost to the relevance of the company's search technology for agents, telling TechCrunch that the system can send back the specific document fragments an agent needs instead of forcing it to read each file in full.

The payoff is largest when paired with more active agents that can operate fluently over the processed data. Box pointed to legal work as an example: lawyers could use the system to locate relevant clauses and precedents more easily. Because Nemotron's embeddings map underlying concepts rather than just keywords, a search can surface related legal ideas even when the documents use entirely different wording.

Hard to copy from scratch

Although the Nemotron models themselves are freely available, the retrieval system depends heavily on Box's pre-existing metadata and permission controls, which makes it difficult to recreate elsewhere. "We're probably packing in five different technologies into one simple API," Levie told TechCrunch. In other words, the open-weights model is only one ingredient; the surrounding infrastructure that knows which user or agent is allowed to see which fragment is where much of the practical value sits.

Why it matters

The announcement is a snapshot of where enterprise software is heading: content platforms are being re-engineered so that agents, not just humans, are the primary consumers of stored documents. Retrieval quality, latency and token efficiency are becoming product-level differentiators for vendors like Box rather than backend details.

It also illustrates the quiet role open-weight models are playing in that shift. While most attention concentrates on frontier models, Nvidia's open-source embedding work is now embedded in a major enterprise platform's core product. And Box's argument that the system cannot simply be rebuilt from the free models alone is a reminder that in enterprise AI, permission controls and accumulated metadata remain the moat — the model is approaching commodity status, while the integration is not.