November 8, 2026 · Knowledge
Your assistant is quoting last year's price list
Retrieval got noticeably better this year. That is not the same as getting the right answer, because the document it found was already wrong — and it cited it beautifully.
There is a sentence that should make anyone running an internal assistant nervous: "According to our documentation."
On Thursday Box announced an agent-first retrieval system built on Nvidia's Nemotron — retrieval that does not just rank documents but acts on what it finds. It is a sensible direction, and an accurate description of where enterprise search is heading: from a list of links to an answer that does something.
It also sharpens a problem that has nothing to do with models. Most companies do not have a retrieval problem. They have an authority problem.
Perfect retrieval of the wrong document
Ask an internal assistant for the minimum order on a part. It answers in two seconds and cites a PDF. The PDF is from 2023, the MOQ changed last spring, and nobody deleted the old file because deleting things feels risky.
The retrieval was flawless. The answer was wrong. A weaker system would have returned five links and let a person notice the date in the footer.
That is the trade you are making as retrieval improves: the failure mode moves from "no answer" to "confident wrong answer", and only one of those costs money.
Three ways it breaks
Duplicates. Three versions of the same spec sheet, all indexed, all equally confident. The model picks one. It has no way to know which one the factory actually works from.
No owner. Nobody is sure who may change the document, so nobody does. Documents rarely decay because people are lazy. They decay because ownership is unclear.
Permissions. An index that inherits access rules is fine. One that does not will eventually surface a salary review in a channel where it does not belong. This is the failure that gets a project cancelled, and it has nothing to do with answer quality.
What actually fixes it
Owner and review date on every document that matters. Not a workflow — two fields: who is accountable, and when it was last checked. Anything past its date gets flagged before it gets quoted.
One canonical document per topic. Everything else gets deleted, not archived.
A test set. Twenty questions with known answers, run quarterly and scored. Not "did the assistant sound convincing" but "did it return the current number".
Citation logging. If an answer came from a document older than twelve months, that is a finding, not a success. Count those and the trend tells you whether the knowledge base is alive.
How to run the test in an afternoon
Pick twenty questions whose answers you know and whose answers have changed at least once. What is the MOQ on this part? Which certificate covers that market? Who signs off on a price exception? Write the current answer next to each one.
Run them through the assistant. Score each answer as right, wrong, or refused, and note which document it cited. The citation matters more than the score: a right answer pulled from a stale file is luck, and luck does not survive the next price change.
Repeat quarterly and track two numbers — accuracy, and the share of answers citing documents older than a year. The second number predicts the first one a quarter ahead.
Why it bites harder in trade
In a business where a wrong HS code means a customs hold and a wrong MOQ means a lost order, "confidently wrong" is not a user-experience problem. It is a cost line with a date attached.
Start with the corpus that carries money: price lists, spec sheets, compliance certificates, and the twenty emails that explain the exceptions to everything above. Two hundred documents, not twenty thousand. If an assistant cannot answer correctly from those, adding volume will not fix it.
The vendor cannot do this for you
Connectors index what exists. They cannot decide which version is true, and they will not tell you that two departments disagree about a price. That judgement stays inside the company, which is inconvenient precisely because it cannot be bought.
The good news is that the work is small and unglamorous: name an owner, delete the duplicates, put a date on the page. The bad news is that no model release will do it for you.
We publish the pricing model, minimum commitment and export-readiness notes for every tool in the library, and we flag it when a vendor changes one without changing the headline price.