Your agent isn't failing. Your retrieval is.

WHAT HAPPENED

Box rebuilt content retrieval for agents on Nvidia Nemotron

Box announced Box Content Retrieval at its developer conference. Instead of reading whole documents, agents get the most relevant chunks. Nvidia's Kari Briski expects faster, more accurate, cheaper queries. Rollout starts early next year.

THE NUMBER

90%

fewer tokens used per query, as Nvidia expects

Also expected: 4x faster answers and 20% fewer errors.

THE REAL PROBLEM IN COMPANIES

What I see in companies: the agent gets blamed for bad search

1. 01The model reads too much, so every answer is slow and costs more
2. 02Nobody tracks answer errors or tokens per query before adding more agents
3. 03Customers use their own words, and the data uses the company's words

I'VE ALREADY SOLVED THIS

A building materials store in southern Brazil: retrieval first, agent second

1. 0113,312 products in the ERP, 136 ways of writing the unit of measure
2. 02245 counter-slang synonyms mapped. The AI never sees the catalog: search hands it 8 results
3. 0365 out of 65 real customer questions found the right product

MY TAKE

Simple that works beats complex that impresses.

Apple is reportedly cutting daily AI spending caps, so tokens are now a real cost. In my case, a simple search answers in milliseconds and the model only sees 8 results. Fix retrieval and measure it before you add agent complexity.