Key points
- Box Content Retrieval is a new feature built on Nvidia's open-weight Nemotron embedding models, announced at Box's developer conference.
- Nvidia's Kari Briski expects speed to quadruple, with errors down 20% and token use down 90%; these are projections, not measured results.
- It reaches Enterprise Advance customers early next year.
- Box says its existing metadata and permission controls make the system hard to rebuild from scratch.
Why it matters for marketers: Content stored in Box becomes cheaper and more precise for agents to search, which affects any team running agent workflows over contracts, briefs or other document libraries.
Box announced a new feature called Box Content Retrieval at its developer conference on Thursday. It builds Nvidia's Nemotron embedding models into the core of the product, so agents can search stored content using metadata the models generate instead of reading whole files.
Box CEO Aaron Levie described the approach this way: the system returns the most relevant chunks of documents rather than making an agent read every file. Nvidia's Kari Briski told TechCrunch she expects answers to arrive four times faster, with 20% fewer errors and 90% fewer tokens consumed. Those figures are her expectations, not independent test results.
For marketers and content teams, the practical point is cost and accuracy. Agents that work over brand guidelines, campaign archives, contracts or approval records burn tokens whenever they scan full documents. Retrieval that targets passages could cut that bill and reduce wrong answers. Box used lawyers as its example, noting that embeddings map concepts rather than keywords, so related material can surface even when wording or language differs.
The Nemotron models are freely available, but TechCrunch reports the system leans on Box's existing metadata and permission controls, which would make it hard to replicate independently. Levie said the offering packs roughly five technologies into one simple API.
The feature goes to Box Enterprise Advance customers early next year. Teams already on that tier should watch for access details and test it on their own content before relying on the claimed gains.
Sources
This brief is our own summary. Read the full reporting at the source.