Internal document search has a particular failure: the most frequently linked page is not the most correct one, and naive relevance keeps surfacing the popular wrong answer.
The internal wiki we were working with had grown for years. Old runbooks sat next to their replacements, and the old ones often had more inbound links because people had bookmarked them long ago. A plain relevance score rewarded that history and kept handing the model outdated steps.
Two adjustments carried most of the improvement. We added a freshness signal so that, among passages of similar relevance, the more recently reviewed one won. And we let document owners mark a page as canonical for a topic, which pinned the right runbook above its own stale copies. Neither of these is machine learning. Both are context decisions about which candidate deserves the space.
The model answered from current runbooks after that. Same retriever, same model. We had just stopped letting the loudest old page win by default.