multi-query retrieval

Multi-query retrieval generates query variations to achieve broader document coverage. **Mechanism**: Original query → LLM generates N alternative phrasings → retrieve with each → merge results (union or RRF). **Why it works**: Single query may miss relevant documents phrased differently. Multiple angles catch variations. Different queries surface different relevant results. **Generation prompts**: "Generate 3 different ways to ask this question", "What related questions might help answer this?", "Rephrase for technical/casual audiences". **Fusion strategies**: Union (all unique results), RRF (ranked fusion), weighted by query similarity to original. **Trade-offs**: N× retrieval cost, increased latency, potential for irrelevant results from poor variations. **Optimization**: Generate queries in parallel, batch embed, efficient deduplication. **Comparison**: Similar to RAG-Fusion which also generates sub-questions and fuses results. **When to use**: Ambiguous queries, exploratory research, broad topics with multiple facets. **Best practices**: Limit to 3-5 variations, validate query quality, monitor result diversity improvement.