How a document assistant decides not to answer
20 November 2026
A knowledge assistant is judged by its wrong answers. The most important thing it can do is recognise when your documents do not cover a question, and say so.
A language model will produce a fluent answer to almost any question. Pointed at company documents, that is a problem: when the documents are silent, a fluent answer is an invented one, and it reads exactly like a real one.
So the behaviour we design first is the refusal. An assistant that says "none of the approved sources covers this" can be trusted with the questions it does answer.
Search first, then decide whether there is anything to say
The assistant never answers from the model's general knowledge. It searches the approved documents, collects the passages that match, and writes only from those passages. Each sentence in the answer points back to the passage it came from.
Before writing, it checks how well the best passages actually match the question. If nothing clears that bar, it stops. It says the sources do not cover the question and, where it can, names who owns the topic.
Respect who is allowed to read what
Search is limited to the documents the person asking is allowed to open. Restricted text never reaches the model, so it cannot leak into an answer. This has to be built into the search itself; filtering the answer afterwards is too late.
Treat unanswered questions as a to-do list
Every question the assistant declines is recorded. Grouped by topic, those questions show which procedures are missing, out of date or impossible to find. The assistant becomes a way to see the gaps in the documents, not only a way to read them.