Query Reformulation Expander

Users write short queries; documents are written in long form. Single-query retrieval often misses passages where the vocabulary differs. Expanding the query into multiple phrasings covers more of the document's vocabulary space, and HyDE retrieves by synthesized-answer similarity — which finds passages that actually answer the question rather than ones that merely restate it.

1. Generate 3 paraphrases of the user query via LLM. 2. Generate one broader and one more-specific variant. 3. Generate a 100-word HyDE passage that pretends to answer the query. 4. Retrieve top-K against each variant and the HyDE passage. 5. Merge via reciprocal rank fusion with a small boost for passages retrieved by multiple variants. 6. Return the top N deduped chunks plus a provenance breakdown showing which variant retrieved each chunk.