Box is adding Nvidia’s Nemotron embedding system to a new feature designed to help AI agents retrieve useful information from documents. Called Box Content Retrieval, the feature is intended to return relevant document passages for a query, rather than making an agent read through each document.
Box says this approach could make queries faster, more accurate and less expensive. The company plans to make the feature available to all Box Enterprise Advance customers, with a rollout early next year.
Retrieving relevant passages for agents
Box Content Retrieval uses embedded metadata created by Nemotron to help agents find information that matches a user’s query or the task at hand. Box CEO Aaron Levie described the change as a way to send back the most relevant chunks from documents instead of having an agent process each document in full.
That distinction matters when an agent has to find a specific detail among large collections of files. If the system can identify relevant passages first, the agent has less material to work through to answer a question or carry out a task.
Nvidia’s VP of generative AI software, Kari Briski, told TechCrunch that she expects queries to be answered four times faster, with 20% fewer errors and 90% fewer tokens used. Those figures are expectations for the system, rather than results reported from a public rollout.
What embeddings add to a search
Nemotron Embed is an Nvidia project offering open-weight embedding models for automated retrieval systems. Such models use large language models to preprocess data so agents can use it more effectively.
Rather than relying only on exact keyword matches, embeddings can map underlying concepts. Box says that can help a search surface related legal themes even when documents use different wording or language.
Box gives lawyers as one example of people who could benefit from active agents working with the feature. They may be able to locate relevant clauses and precedents more easily when a search can connect related ideas across documents.
The feature depends on Box’s existing systems
The Nemotron models are freely available, but Box Content Retrieval also relies heavily on Box’s existing metadata and permission controls. Those pieces are part of how the service organizes content and manages access, and Box says they would make the system challenging to recreate from scratch.
Levie told TechCrunch that the product combines “five different technologies into one simple API.” The embedding model is therefore one part of a broader retrieval feature built around Box’s own content and controls.
For customers, the planned availability is specific: Box says the feature will roll out early next year to all Box Enterprise Advance customers. The announcement describes its intended benefits and use cases, while the expected performance figures come from Nvidia’s Briski.