Box integrates Nvidia’s Nemotron embeddings to speed up agent-driven content retrieval
Box’s new Content Retrieval feature uses Nvidia’s Nemotron embeddings to cut query time by 75%, errors by 20%, and token use by 90% for Enterprise Advance customers.
1 source · cross-referenced
- Box will integrate Nvidia’s Nemotron embedding models into its Content Retrieval feature for Enterprise Advance customers.
- The integration aims to make agent-driven queries four times faster, reduce errors by 20%, and cut token usage by 90%.
- Nemotron embeddings map underlying concepts, enabling retrieval of relevant content even when wording differs.
- The feature relies on Box’s existing metadata and permission controls, making recreation from scratch difficult.
- Rollout begins in early 2027 for all Enterprise Advance customers.
Box announced it will integrate Nvidia’s Nemotron embedding models into a new Content Retrieval feature for its Enterprise Advance customers. The feature is designed to let agents answer queries more effectively by using Nemotron-generated embedded metadata. Box claims the integration will make queries four times faster, reduce errors by 20%, and decrease token usage by 90%.
Nvidia’s VP of generative AI software, Kari Briski, told TechCrunch the speed and efficiency gains are expected to materialize when Nemotron embeddings are used in Box’s retrieval pipeline. The system will be available to all Enterprise Advance customers, with a rollout planned for early 2027.
Box CEO Aaron Levie said the feature improves search relevance for agents by sending back only the most relevant document chunks, rather than requiring agents to read entire documents. Nemotron embeddings go beyond keyword matching to map underlying concepts, enabling retrieval of related content even when different wording is used.
The integration relies on Nemotron’s open-weight embedding models, which Nvidia released in July and which topped the RTEB benchmark for retrieval quality. While the models are freely available, the system also depends on Box’s pre-existing metadata and permission controls, making it difficult to recreate from scratch.
Levie described the feature as combining multiple technologies into a single API, emphasizing its utility for domains like legal services where agents need to find relevant clauses and precedents quickly.
- Nov 11, 2026 · Hugging Face
Hugging Face proposes token pooling-aware training to improve compressed multi-vector retrieval
Trust79 - Nov 11, 2026 · TechCrunch — AI
Fractile raises $720M Series C to develop AI chips and compiler stack
Trust79 - Nov 11, 2026 · The Verge — AI
Google’s Tabletop AI experiment lets users build custom collaborative apps via waitlist
Trust79