Query generation transforms one ambiguous user input into multiple retrieval queries. In RAG-Fusion, it widens coverage by targeting different facets or sub-intents instead of merely producing near-identical rephrasings. section 8.2 section 8.1
The generated queries run in parallel against the retrieval system. Their results provide a broader candidate set, which a later fusion step filters into a more complete context for answer generation. section 8.4 section 8.1
A useful generator decomposes an apparently simple request into distinct information needs:
- “What is the best laptop for college?” can produce queries about durability, portability, budget, and battery life. section 8.2
- Simple rephrasings often retrieve overlapping documents because they preserve the same average intent. section 8.2
- Increasing query breadth differs from increasing one query’s depth: several queries with small k seek wider coverage before fusion restores precision. section 8.1
Prompt-driven generation
A prompt can ask an LLM to generate N alternative search queries, each targeting a different facet, sub-intent, or interpretation, with one query per line. The approach requires no specialized rewriting model, although dedicated models remain an option when generation becomes a larger concern. section 8.2
You are a search assistant. The user has asked: "{question}". Generate {N} alternative search queries that target different facets, sub-intents, or interpretations of this question. Each query should explore a distinct angle so that together they cover the full intent. Return one query per line, no numbering.
Use the supplied prompt template to generate distinct search perspectives rather than surface-level paraphrases. section 8.2
