Enhancing AI Agent Retrieval
Box has announced the integration of Nvidia’s Nemotron embedding system into its core product architecture. This update aims to refine how AI agents interact with stored content.
The new feature, Box Content Retrieval, utilizes embedded metadata generated by Nemotron to allow AI agents to process and answer queries with greater efficiency.
Performance and Efficiency Gains
According to Nvidia, the implementation is expected to result in queries being answered four times faster, with a 20% reduction in errors and a 90% decrease in token usage compared to previous methods.
The system is scheduled for release early next year for all Box Enterprise Advance customers.
Box CEO Aaron Levie noted that the technology improves search relevance by allowing agents to receive specific, highly relevant document chunks rather than requiring the system to process entire files for every task.
Leveraging Nemotron Embed
The feature implements Nvidia’s Nemotron Embed project, which provides open-weight embedding models designed to optimize retrieval systems. These models pre-process data to enhance agent utility, a field that has recently seen increased focus from Nvidia, including the July release of models that performed well on the RTEB retrieval benchmark.
Conceptual Search Capabilities
By moving beyond simple keyword matching to map underlying concepts, the system allows for more nuanced retrieval. For instance, legal professionals could use the tool to identify relevant clauses and precedents even when the terminology varies between documents.
While the underlying Nemotron models are open, the system integrates with Box’s existing metadata and permission frameworks, creating a unified API that simplifies complex retrieval workflows.
