Refusal is an outcome
The most underrated path in a retrieval system is the one that does not answer.
Most retrieval demos have one exit: an answer. Ask something the corpus cannot support and the system does what it was built to do. It finds the nearest chunks, however far away, and writes fluent prose over them.
The fix isn't a better prompt. It is a bigger set of outcomes. When I built a retrieval agent over research papers, the first node decides what kind of turn this is: retrieve, call a tool, ask a clarifying question, refuse, or simply chat. Refusing and clarifying are not error states. They are paths with their own tests.
That changes how you evaluate. A benchmark of only answerable questions rewards a system that never says no. So the evaluation includes questions that should be refused, questions that are ambiguous, and greetings that should never touch the corpus. If the “no” paths aren't in the test set, they will quietly rot.
A system that can't say “I can't support that” will eventually say something false. Give it the sentence, and then make sure it's tested.