AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Nov 5, 2026

Box used its developer conference on Thursday to unveil Box Content Retrieval, a feature that puts Nvidia's Nemotron embedding system at the center of how the company's product answers queries on behalf of AI agents. The company says the result for users is faster, more accurate and cheaper retrieval. It will be offered to all Enterprise Advance customers, with a rollout planned for early next year.

The performance figures come from Nvidia rather than from Box's own testing. Kari Briski, Nvidia's vice president of generative AI software, told TechCrunch she expects queries to be answered four times faster, with 20% fewer errors and 90% fewer tokens consumed. Those three numbers describe different things: latency, accuracy, and the volume of model tokens burned to produce an answer. The token reduction is arguably the most consequential for cost, since inference bills scale with token counts.

The mechanism Box describes is retrieval over embedded metadata. Rather than an agent reading through whole documents to find what it needs, the system uses Nemotron-generated embeddings to locate the most relevant passages and returns just those chunks. Box CEO Aaron Levie framed the benefit as relevance for agents, saying the product can send back the chunks most relevant to a query or task instead of making the agent read each document.

Embeddings are vector representations that place text with similar meaning near each other in a mathematical space, which lets a search system match on concepts rather than literal keywords. That distinction is central to Box's pitch. Because the representations map underlying concepts, a query can surface related material even when the wording or the language differs. Box's stated example is legal work: lawyers could find relevant clauses and precedents more easily, including related legal themes expressed in different terms. This is the kind of task where keyword search tends to fail, since the same idea can appear under many phrasings.

The underlying technology is an implementation of Nvidia's Nemotron Embed project, which supplies open-weight embedding models aimed at improving automated retrieval. Nvidia has made this area a focus of its open-model work even though, as the reporting notes, major labs have often overlooked embedding models in favor of generative ones. Nvidia's most recent Nemotron release in July topped the RTEB benchmark, which measures retrieval quality.

Open weights matter here in a practical sense: the models themselves are freely available, so the embedding layer is not the moat. Box argues the harder part to replicate is everything around it. The system leans heavily on Box's existing metadata and permission controls, and Levie said the company is effectively packing five different technologies into one API. For a developer reading this, that suggests the integration and governance layer, not the model, is where the differentiation sits. Rebuilding something comparable from scratch would mean combining retrieval, metadata handling, access control and query orchestration, and getting them to behave consistently.

Permission controls deserve particular attention in enterprise retrieval. An agent that returns chunks from documents must respect who is allowed to see those documents, and retrieval systems that ignore access boundaries can leak content across teams. Box is positioning its existing controls as part of the value, which implies the retrieval layer is scoped to each user's entitlements.

A likely trade-off is that the headline numbers are expectations tied to a specific configuration: Nemotron embeddings inside Box's stack, over Box's metadata. They may not transfer to other deployments or other corpora. Retrieval quality also depends heavily on how documents are chunked and how metadata is maintained, so results will vary with how well a customer's content is organized. The early-next-year timing means these figures have not yet been validated by customers at scale.

Why it matters: enterprises building agents over large document stores face a choice between assembling retrieval themselves and buying it inside a platform they already use. Box is betting that permission-aware retrieval over existing metadata is the harder half of the problem, and that pairing Nvidia's open embedding models with its own controls is enough to win that bet. If the projected gains hold, teams could see meaningfully lower token costs and fewer failed lookups per agent task; that outcome is an expectation, not a demonstrated result. Developers weighing in-house retrieval stacks should watch how the feature handles access control and chunk quality once it ships.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

The system can supposedly answer queries four times faster, with 20% fewer errors and 90% fewer tokens used.