What is query transformation in RAG and why does it help?

I see query transformation as a query-rewrite step that runs before retrieval. The system takes the user's question and produces one or more search queries that are more likely to match the indexed text. This is useful because people and documents often describe the same concept in different language, so a semantically relevant document may not look close to the original wording.

Depending on the case, the rewrite can resolve pronouns from earlier chat turns, expand an acronym, or split a multi-part question into separate subqueries. For example, if someone asks what it supports after describing a company product, I would substitute the relevant product name before searching. Another technique is HyDE. I generate a hypothetical answer, embed that answer rather than the question, and use the resulting vector to find supporting passages. The important point is that each transformed query still retrieves evidence that I can inspect before answering.

Transformation is not automatically better, though. Model-based transformation can add calls and latency and may introduce errors. I would compare the transformed and original queries on a fixed retrieval set, looking at relevant-document recall and answer quality rather than assuming a gain. I would use it for ambiguous, conversational, or multi-part questions, and keep the original query as a fallback. That lets me pay the extra latency only when the measured recall improvement justifies it.