query rewriting

People ask messy, short, context-dependent questions. "What about its battery?" means nothing to a retriever on its own — it does not know what "it" is. Query rewriting reformulates the user's raw question into a cleaner, self-contained search query before retrieval runs, so the search actually goes looking for the right thing.

An LLM commonly does the rewriting: it resolves pronouns using the earlier conversation, adds implied terms, fixes typos, or splits one compound question into several separate searches. Related tricks generate multiple paraphrased queries, or even draft a hypothetical answer and embed that to search with. Because the quality of retrieval caps the quality of the whole pipeline — the model can only ground on what was fetched — sharpening the query is often the single highest-leverage improvement you can make.