Unit 04.01: Expanding, narrowing, and splitting a question
Rewriting is not one operation. It is three, they fix different problems, and applying the wrong one makes retrieval worse rather than better.
Expand, narrow, split
Expanding adds vocabulary, for the case where the user's words do not appear in the corpus. Narrowing adds constraints, for the case where the corpus returns too much. Splitting divides a compound question into the separate retrievals it actually needs.
The example below applies all three to one question and prints what each produces.
expand : 3 queries
'refund window'
'refund period'
'return policy timeframe'
narrow : 1 query
'refund window individual plans'
split : 2 queries
'what is the refund window'
'does the refund window apply to enterprise'
A compound question retrieved as one query answers half of itself.
The question - "What is the refund window and does it apply to enterprise?" - is really two questions. Expanding it or narrowing it still produces a single retrieval, and a single retrieval can only satisfy one half.
Splitting is therefore the correct operation here, and it is the one people skip, because it changes the pipeline shape: two retrievals, two result sets, and an answer that has to combine them and say which part came from where.
The mistake this prevents
The mistake is applying expansion by default because it is the easiest to implement. Expansion on an already-well-matched query adds noise terms that pull retrieval away from the right chunk - the failure the last unit of this module demonstrates. Diagnose which of the three problems you have before choosing.
Takeaway
Match the rewrite to the failure: expand for vocabulary mismatch, narrow for an over-broad corpus, split for compound questions. A compound question retrieved as one query answers half of itself and says nothing about the other half.
