Query Rewriting & Decomposition

Query Rewriting & Sub-Query Decomposition is a core Pre-Retrieval RAG optimization technique that transforms raw user prompts into structured retrieval intents; using specialized prompts or fine-tuned LLMs, it performs coreference resolution (e.g., resolving 'its latest report' into 'Apple Inc. 2026 Q2 Financial Report'), synonym query expansion, and breaks complex multi-hop queries down into parallel or sequential atomic sub-queries that query the index independently before consolidating.

Multi-turn conversational RAG, multi-entity comparative analysis (e.g., 'Compare Nvidia and Qualcomm newest chip efficiency and price'), and ambiguous long-tail questions.

Raw user questions are colloquially ambiguous, pronoun-laden, and multi-faceted; searching composite queries directly yields poor vector similarity and misses cross-document facts; decomposing into atomic queries and aggregating results unlocks multi-hop reasoning.

Describe the prompt pattern for performing conversational coreference resolution and synthesizing standalone queries from chat history?

Analyze latency vs accuracy trade-offs between parallel sub-query dispatch and sequential DAG step-by-step query execution?

How does Multi-Query Retrieval generate 3-5 diverse synonym query variations to maximize overall retrieval recall?

How does Step-Back Prompting generate high-level abstracted concept queries to augment RAG background retrieval?

How to constrain query rewriting LLM prompts to prevent hallucinated over-generalization drifting from the user's core intent?