Query Rewriting

Query rewriting transforms a user request into one or more retrieval queries. It is common in RAG systems because the user’s natural-language question is not always the best search query.

A rewriting function maps:

$$R(q,h,s)\to \{q_1^{\prime },\dots ,q_k^{\prime }\}$$

where $q$ is the current query, $h$ is conversation history, and $s$ is task state.

Rewrite Types

• Clarifying rewrite: make implicit references explicit.
• Decomposition: split a complex question into subqueries.
• Expansion: add synonyms, acronyms, or related concepts.
• Constraint extraction: turn a request into filters such as date, source, domain, modality, or entity.
• HyDE-style rewrite: generate a hypothetical answer or passage, then retrieve against it.

Risks

Query rewriting can improve recall but can also drift from the user’s intent:

$$\mathrm{drift}(q,q^{\prime })=1-\mathrm{sim}(\mathrm{intent}(q),\mathrm{intent}(q^{\prime }))$$

The rewrite should be treated as a retrieval aid, not as a new task definition.

Checks

• Does the rewrite preserve the original task and constraints?
• Are filters extracted from evidence or invented?
• Is the rewritten query visible in logs or artifacts for debugging?
• Are multiple rewrites merged without flooding the context window?
• Is retrieval measured before and after rewriting?
• Can the system fall back to the original query?

Related

• Retrieval-augmented generation
• Hybrid retrieval
• Context packing
• Retrieval
• Question answering
• Context engineering