Better Together: Complementary Query Rewriting Under a Strong RAG Baseline

Rephrasing a question several ways can help an AI system find supporting documents, but the extra searches are useful only if they recover information the original search misses. This study's abstract reports that individual rewriting methods were, at best, competitive with an already strong search baseline.

Combining methods helped on enterprise documents because different methods missed different questions, yet the same approach hurt performance on a dataset of ambiguous questions. A simulated router that requested rewrites only when the original search's confidence was low recovered part of the enterprise improvement at lower cost. The supported direction is selective use of complementary searches, with the choice tested against the actual question mix.

“Our headline result is that rewriting alone is at best competitive with a strong baseline, but combining methods yields outsized gains because different strategies fail on different questions.”

“In short: treat query rewriting as a complementary coverage source, applied through cost-aware routing, not as a standalone replacement for a strong baseline.”