Rendered Source Note

Query Rewriting for Retrieval-Augmented Large Language Models (Rewrite-Retrieve-Read)

Generated HTML view. Markdown remains canonical.

Query Rewriting for Retrieval-Augmented Large Language Models (Rewrite-Retrieve-Read)

Type: paper Tier: 2 (Foundational Paper) URL: https://arxiv.org/abs/2305.14283 Accessed: 2026-06-09 Authors: Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, Nan Duan Year: 2023 (EMNLP 2023) arXiv: 2305.14283 — https://arxiv.org/abs/2305.14283

Primary source for Elective 05 — Advanced RAG / Query Engineering (query rewriting, M1). Faithful summary; arXiv PDF is canonical.


Core idea

Standard RAG is retrieve-then-read: take the user's question as-is, retrieve, answer. But "there is inevitably a gap between the input text and the needed knowledge in retrieval" — the way a user phrases a question is often not the best query for the index. This paper inserts a step: Rewrite → Retrieve → Read. Reformulate the question into a better retrieval query first.

Approach

Two forms:

  1. LLM rewriter — prompt an LLM to produce one (or several) search queries from the question.
  2. Trainable rewriter — a small LM trained with **reinforcement learning from the reader's

feedback** (the black-box LLM reader's answer quality is the reward signal).

Result

Consistent improvement over retrieve-then-read across open-domain and multiple-choice QA — the rewrite closes the question↔query gap before retrieval ever runs.

Why This Source Matters

It is the first and most general query transform: the user's words are an input to be engineered, not a fixed query. It pairs with HyDE (rewrite the query text vs. embed a hypothetical answer) and multi-hop decomposition (one query vs. several). And like all of them, it can drift — a rewrite can wander off the user's intent — so it is measured against the eval, not assumed good.

Key Claims

Relevant To

Known issues / cautions