Perspective shifting formulates searches from distinct facets, sub-intents, or interpretations of one information need. In multi-query retrieval, an LLM expands the original question before retrieval, helping the system cover complex intent instead of relying on one semantic guess. section 8.0 section 8.2
Use perspective shifting to improve coverage without treating more queries as automatically better:
- Target different needs rather than producing simple rephrasings, which often retrieve nearly identical documents. section 8.2
- Run the variants in parallel, then merge their ranked results with Reciprocal Rank Fusion, which uses rank rather than incompatible retrieval scores. section 8.0 section 8.7
- Start with three to five variants and evaluate latency, cost, coverage, and noise; additional queries can reduce quality. section 8.7
Winter jacket search
A standard search for “warm winter jacket” may return several near-duplicate warm-jacket documents. Perspective shifting generates queries about down parkas for extreme cold, lightweight insulated fleeces, waterproof ski jackets, and wool coats. Fused retrieval then surfaces distinct product categories, moving from narrow relevance matching toward discovery. section 8.5
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 this prompt pattern to instruct an LLM to produce one query per distinct facet, sub-intent, or interpretation. section 8.2
