Box Context Retrieval: Nemotron-powered search for AI agents

Box announced Box Context Retrieval at BoxWorks 2026, a retrieval engine built on NVIDIA's open Nemotron Retriever embedding models. Box says it feeds AI agents only the document chunks they need, with permissions applied, instead of making the agent read whole files. The headline claims are big: up to 90% fewer tokens, 20% better search accuracy and 4x faster answers. The catch is in the wording. Box says customers "will have the potential" to get those gains, and the product isn't generally available until early next year.

What actually happened

Per Box's announcement on November 5:

• The engine: Box Context Retrieval, powered by NVIDIA Nemotron Retriever embedding models.
• Features: permission-aware retrieval across millions of files, hybrid keyword plus semantic search, and indexing of scanned documents, images, audio and video.
• Claimed gains: 20% better accuracy (Recall@5), 90% fewer tokens and 4x faster question answering. All three are framed as potential, not measured customer results.
• Availability: generally available to Enterprise Advanced customers early next year. Nemotron models, starting with 3 Ultra, are in Box AI for those customers now.
• Pricing: not announced.

TechCrunch quotes CEO Aaron Levie describing it as several technologies packed "into one simple API." The Nemotron models are free and open-weight. What Box sells is the part that's hard to copy: years of file metadata and the permissions already set on every folder.

Why it matters for your business

That 90% token claim is the real story, and it holds up even without Box. Most of what an AI agent costs when it works over your documents comes from stuffing whole files into the context window. Good retrieval, meaning chunked, ranked and filtered by who's allowed to see what, is the cheapest way to make document agents both cheaper and safer.

Permissions are the part small teams skip. A homemade RAG pipeline that indexes a shared drive with one service account will happily show the intern a payroll spreadsheet. Box's pitch is that its retrieval respects the same access rules people already have. Whatever you build on, make sure your retrieval layer does the same.

What we'd do now:

• If you're on Box Enterprise Advanced, test the Nemotron models in Box AI now and benchmark Context Retrieval when it ships. Measure tokens per answer yourself.
• If you're not, don't upgrade a plan for a feature with no date and no price.
• Either way, check whether your current document agent filters by user permissions before retrieval, not after.

Sources: Box — Box Partners With NVIDIA to Deliver the Trusted Content Layer for Enterprise AI, TechCrunch.