At 03:07, the page says that answers over the logistics support archive have started citing the same incident repeatedly. The on-call opens one trace and sees three PDF tickets about a delayed container, all admitted as separate evidence even though two are near copies created by an email export. The least complex useful response is a similarity gate with three outcomes: accept a clearly new ticket, suppress a clear duplicate, and send the uncertain middle to review. Do not turn one similarity threshold into an automatic delete button.

TL;DR: normalize and embed each ticket with the same pipeline, search only within an operationally meaningful candidate set, and treat similarity as a ranking signal rather than truth. Keep both sides of every decision, the model and preprocessing version, the score, the chosen boundary, and the reviewer outcome. Page on user-visible retrieval damage or sustained decision drift, not on a lone high score.

This matters before generation.…