query rewriting

Query rewriting transforms user queries to better match document format before retrieval. **Problem**: Users ask natural questions but documents are written in different style. "What causes headaches?" vs document "Headache etiology includes...". **Techniques**: **LLM rewriting**: Use model to rephrase query in document style, expand abbreviations, add context. **Query expansion**: Add synonyms, related terms, domain vocabulary. **Decomposition**: Break complex query into sub-queries. **Correction**: Fix typos, normalize terminology. **HyDE approach**: Generate hypothetical answer, use that for retrieval. **Multi-query**: Generate variants covering different phrasings. **Implementation**: Query → LLM rewriter → enhanced query → retrieval. **Prompting**: "Rewrite this question as it might appear in a technical document" or "Generate search terms for:". **Evaluation**: Compare retrieval metrics (recall@k, MRR) before/after rewriting. **Trade-offs**: Adds latency (LLM call), may introduce errors, cost per query. **When essential**: Complex questions, domain mismatch between users and documents, ambiguous queries. Significantly improves RAG retrieval quality.