Box taps Nvidia's Nemotron for faster, cheaper AI agent content retrieval

What Box Content Retrieval does

At its BoxWorks 2026 developer conference on Thursday, Box showed off a new feature that builds Nvidia's Nemotron embedding system into the core of its product. Called Box Content Retrieval, the feature allows AI agents to answer queries by relying on embedded metadata created by Nemotron, rather than reading through every document.

Instead of an agent having to read through each document, Box CEO Aaron Levie told TechCrunch, the system 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 result, according to Nvidia's vice president of generative AI software Kari Briski, is queries answered four times faster, with 20 percent fewer errors and 90 percent fewer tokens used.

How Nemotron embedding works

Nemotron Embed is Nvidia's open-weight embedding model project, which uses LLMs to pre-process data so it can be better used by agents. The approach goes beyond simple keyword matching to map underlying concepts, so a search tool can surface related themes even when different wording or language is used. Nvidia's Nemotron embedding models have been a focus of the company's open-source model efforts.

The July 2026 Nemotron release topped the RTEB benchmark, which measures how well retrieval systems surface relevant content. That benchmark performance is why Box chose the models for its retrieval layer, rather than a general-purpose embedding model from a cloud provider.

While the Nemotron models themselves are freely available under open licenses, Box Content Retrieval also relies heavily on Box's pre-existing metadata layer and permission controls, making it difficult for competitors to recreate the system from scratch. We're probably packing in five different technologies into one simple API, Levie told TechCrunch.

Why retrieval matters for AI agents

Embedding-based retrieval is becoming critical infrastructure for enterprise AI, where agents need to find relevant information across thousands of documents without paying the cost of sending entire files to a frontier model. The 90 percent token reduction Nvidia claims represents a significant cost saving for enterprises running AI agents at scale.

Box gave the example of lawyers who will be able to find relevant clauses and precedents more easily under the new system. Because Nemotron's embeddings capture conceptual relationships, a lawyer searching for indemnification obligations would also surface clauses about liability coverage even when the exact terms differ.

The system follows a broader trend of embedding models becoming a competitive battleground in enterprise AI. A growing number of providers offer specialized embedding services optimized for different retrieval tasks, and retrieval quality has become a headline metric in model benchmarks.

Box says the retrieval layer also improves accuracy for non-agent use cases like search and regulatory discovery, where finding every relevant document matters more than speed. Permission-scoped retrieval means results are filtered by what the user is allowed to see, an important compliance feature for regulated industries that Box says is baked into the product rather than added on.

Availability and rollout

Box Content Retrieval will be available to all of Box's Enterprise Advanced customers, rolling out early next year. The system is designed to work alongside active agents, with the pre-processed embeddings making retrieval fast enough for real-time use.

The announcement was part of a broader BoxWorks 2026 event where Box also unveiled other AI-focused features including an AI workflow builder. The expanded Nvidia partnership signals deepening ties between the two companies as enterprise content management increasingly shifts toward AI-native workflows.

For Nvidia, the Box deal represents another distribution channel for its Nemotron family, following integrations with other enterprise platforms. The company has been pushing Nemotron embedding models as foundational infrastructure for enterprise AI retrieval, competing with both open-source alternatives and proprietary embedding services from cloud providers.

The embedding model race

Box's choice of Nemotron reflects a broader shift in how enterprises build AI search and agent infrastructure. For years, vector databases and embedding models were an afterthought — teams bolted on whatever embedder their database vendor shipped. That is changing as retrieval quality becomes the difference between an agent that answers correctly and one that hallucinates from the wrong documents.

Nvidia's Nemotron Embed line has pushed into this market with open weights and aggressive benchmark results, and the company has been signing up enterprise platforms to embed its models. Box becomes the latest enterprise content platform to build retrieval on top of Nemotron, giving the chipmaker a software foothold in enterprise content management.

The economics matter too. Embedding pre-processing happens once and is relatively cheap, but the savings compound on every query. A 90 percent token reduction on agent queries can cut inference spend dramatically for enterprises running thousands of agent calls a day, which is the pitch Box and Nvidia are making to CIOs.