The latest in AI, every dayAI News
← AI News

November 5, 2026 · TechCrunch (Russell Brandom)

Box Taps Nvidia's Nemotron for New Agent-First Content Retrieval System

My take: Box unveiled "Box Content Retrieval," a feature built on Nvidia's Nemotron embedding models, so an AI agent receives only the relevant chunk of a document instead of having to read the whole file. Box CEO Aaron Levie put it this way: the agent no longer processes an entire file to find one clause, it processes just the fragment that matters. Nvidia, for its part, estimates (its own figure, not an independently verified one) roughly 4 times faster queries, 20% fewer errors, and 90% fewer tokens used. The feature rolls out to all Box Enterprise Advance customers in early 2027.

That number is worth reading with a bit of care: it comes from Nvidia VP Kari Briski, about a product built on her own company's technology. I don't doubt the approach itself (searching by concept instead of keyword, and pulling only the needed fragment) is a real improvement; that's how any retrieval system for agents should work. But it's worth waiting for numbers from actual customers before treating those figures as settled.

What matters beyond the exact percentage is the problem it solves: today, many AI agents spend time and token budget reading entire documents to find a single answer. For any business handling contracts, manuals, or long files (lawyers searching for clauses, compliance teams reviewing policies), a system that understands content by concept, not just keyword, can be the difference between a useful agent and one that costs more than it saves.

If your business already uses AI to search your own documents, do you know how much of that token cost goes toward reading pages that never had the answer?

Read at the source: TechCrunch (Russell Brandom) ↗

Want to use these tools? See the unbiased reviews or back to the news.