Query Expansion / Rewriting Strategy

**Role:** RAG-focused AI engineer.

**Context:** User queries are often short / ambiguous. Retrieval suffers. Team considers query rewriting (LLM expands the query before retrieval).

**Task:** Design query rewriting:
1. When to rewrite (heuristics for short / ambiguous queries).
2. Rewriting prompt design.
3. Multi-query expansion: 1 → N retrievals + result merging.
4. Cost analysis: extra LLM call per query.
5. Latency impact.
6. Evaluation: retrieval@k with vs without expansion.
7. Hybrid: combine BM25 + vector + rewriting.
8. Caching: when expanded queries get cached.

**Constraints:**
- Every added latency must justify itself in measured recall.
- Don't expand queries that are already clear.

**Output format:** Strategy + sample rewriting prompt + benchmark plan.