You already know: sentence embeddings and bi-encoder versus cross-encoder trade-offs (D4 and D7), approximate neighbour search (the visual search case), calibration (D3) and, from the autocomplete case, that what the system says publicly needs strict rules. This case adds matching by intent in a community: two questions are duplicates when one answer serves both, which is different from being similar; mistakes have social cost; and the labels come from moderators whose decisions are imperfect. The design questions: what a duplicate is, how to represent intent, how strong an action each confidence level deserves, and how to learn from human decisions.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud.
1. The prompt, and the questions that change the design
A mid-level answer: "embed questions and close any new question whose cosine similarity to an old one is above 0.9". A staff candidate first asks what counts as a duplicate, when the system acts (while typing, after posting), what labels exist and what each mistake costs. Those answers produce an intent-level definition with hard negatives, tiered actions from suggestions to merges, moderator closures and reopens as labels, and reversible actions with high-precision thresholds.
Assumptions you state: