multi-query retrieval

**Multi-query retrieval** is the **strategy of generating multiple query variants for one information need and retrieving with each to improve coverage** - it increases recall by exploring different semantic angles. **What Is Multi-query retrieval?** - **Definition**: Retrieval approach that decomposes or reformulates a query into diverse sub-queries. - **Variant Sources**: LLM paraphrases, subtopic prompts, intent facets, or domain-specific rewrites. - **Fusion Step**: Results are merged, deduplicated, and reranked into a unified candidate list. - **Pipeline Role**: Improves first-stage evidence discovery before generation. **Why Multi-query retrieval Matters** - **Recall Expansion**: Captures documents missed by single-query lexical or semantic mismatch. - **Complex Question Support**: Better handles broad or multi-faceted user requests. - **Robustness Gain**: Reduces dependence on one imperfect query phrasing. - **RAG Reliability**: More complete evidence sets improve grounded answer quality. - **Tradeoff**: Increases retrieval compute and requires stronger dedup and ranking controls. **How It Is Used in Practice** - **Variant Budgeting**: Limit number of generated queries by latency constraints. - **Result Fusion**: Apply reciprocal rank fusion or learned merging with duplicate suppression. - **Adaptive Triggering**: Use multi-query only when baseline retrieval confidence is low. Multi-query retrieval is **a practical coverage-boosting technique in RAG pipelines** - diversified query generation plus robust fusion often yields meaningful improvements on difficult information needs.