Box Bakes Nvidia's Nemotron Into Its Core for Agent-First Retrieval

Box says its new Nvidia-powered retrieval system will answer queries four times faster, with 20% fewer errors and 90% fewer tokens used. Nvidia VP of Generative AI Software Kari Briski shared those figures with TechCrunch at Box's developer conference on Thursday, where the company announced Box Content Retrieval — a feature that embeds Nvidia's Nemotron embedding system into the core of its product.

The deal puts open-weight embedding models, a category frontier labs have largely ignored, at the center of one of the largest enterprise content platforms. Box is betting that retrieval quality, not just raw model power, decides whether agents actually work inside companies.

What does Box Content Retrieval actually do?

The feature lets AI agents answer queries by relying on embedded metadata created by Nemotron, rather than reading entire documents. Box CEO Aaron Levie said the change is fundamental to how the platform serves agentic workflows.

"This dramatically boosts both the relevance of our search technology for agents," Levie told TechCrunch. "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 practical result for users: faster, more accurate and cheaper queries. The system rolls out early next year and will be available to all Enterprise Advance customers.

Why Nemotron?

The feature implements Nvidia's Nemotron Embed project, which offers open-weight embedding models designed to improve automated retrieval systems. Embedding models use LLMs to pre-process data so it works better with agents — a technical corner that major AI labs have often skipped, but one Nvidia has made a priority within its open-source model program.

The bet is paying off on benchmarks. Nvidia's most recent Nemotron release in July ranked #1 on the RTEB benchmark, which measures retrieval quality. That ranking gives Box a defensible technical foundation rather than a proprietary black box.

Where the advantage shows up

Box says the system works best when paired with more active agents that can operate fluently over the Nemotron-processed data. The company pointed to lawyers as a concrete use case: finding relevant clauses and precedents becomes easier because Nemotron's embeddings map underlying concepts rather than simple keywords.

That distinction matters. A search tool built on conceptual embeddings can surface related legal concepts even when the documents use entirely different language — something keyword search can't do.

Can competitors copy it?

Not easily, and that's the strategic point. The Nemotron models themselves are freely available to anyone. But the retrieval system also leans heavily on Box's pre-existing metadata and permission controls, which took the company years to build and layer across its customer base. Replicating the result from scratch would mean rebuilding that infrastructure too.

Levie framed the integration as a compression of multiple technologies into a single interface. "We're probably packing in five different technologies into one simple API," he told TechCrunch.

The move signals where enterprise software competition is heading: platforms that already own structured access to corporate data can pair with open models to deliver agentic features that standalone AI startups struggle to match. For Box, the early-2026 rollout to Enterprise Advance customers will be the first test of whether 4x speed and 90% token savings translate into renewals.