所属模块：
M7 · 检索、排序与推荐系统 (Retrieval, Ranking & RecSys)| 专题分类：混合检索与融合 (Hybrid Retrieval & RRF Fusion)| 难度等级：Hard
一、核心一句话结论 (One-Sentence Summary)
改写（补全指代/纠错/分解）与扩展（加同义词/相关词）可提升召回；但可能引入噪声（查询漂移）。
Query rewriting refines, disambiguates, and decomposes user queries, while query expansion appends synonyms and related terms; techniques range from classical pseudo-relevance feedback (PRF) to LLM-generated hypothetical documents (HyDE).
二、核心考点要义 (Key Insights)
- 📌 改写：补全上下文、纠错、分解多跳问题、调整表述
- 📌 扩展：加同义词、相关词、上位词（提升召回）
- 📌 风险：查询漂移（扩展偏了导致噪声）、成本（额外 LLM 调用）
English Insights:
– Rewriting vs. expansion: Rewriting transforms flawed queries (spelling correction, coreference resolution, multi-hop decomposition); expansion adds relevant vocabulary.
– Query drift hazard: Over-aggressive expansion introduces irrelevant topical associations, degrading precision.
– Generative methods (HyDE): LLMs generate hypothetical answer passages that dense retrievers use as rich semantic search vectors.
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{rewrite}: qto q’;qquad text{expand}: qto q+Delta;qquad text{risk}: text{query drift}$$
数学机理：查询改写与扩展——(1) 改写（rewriting）——把查询变换为更’可检索’的形式：(a) 上下文补全（多轮对话中’它’指什么）；(b) 拼写/语法纠正；(c) 多跳分解（’A 的创始人的母校’ → 拆成’A 的创始人’ + ‘其母校’）；(d) 意图澄清（’苹果’ → ‘苹果公司’ 或 ‘苹果水果’）；(e) 表述规范化（口语 → 书面）。(2) 扩展（expansion）——增加词（而非替换）：(a) 同义词（’汽车’ → ‘汽车 轿车 车辆’）；(b) 相关词/共现词；(c) 上位词/下位词（’水果’ → ‘水果 苹果 香蕉’）；(d) 翻译（跨语言扩展）；(e) 伪相关反馈（PRF/RM3）——先用原查询检索、假设 top-k 相关、从中提取扩展词。收益——提升召回（尤其对’短查询/词汇失配’）；代价——(a) 查询漂移（query drift）——扩展词偏了（PRF 尤其明显：若首次检索差则越走越偏）；(b) 精度下降（引入噪声）；(c) 成本（额外 LLM 调用 / 额外检索）；(d) 延迟（多一次往返）。什么时候该做——(a) 短查询/歧义查询——收益大；(b) 多轮对话——必须补全上下文；(c) 多跳问题——必须分解；(d) 长且明确的查询——不该做（已足够明确，扩展只会引入噪声）；(e) 精确匹配需求（如’查找 ERR_403’）——不该扩展（会破坏精确性）。实现方式——(a) 规则（同义词表、词典）；(b) 统计（PRF/RM3、词共现）；(c) 模型（用 LLM 改写/扩展——当前主流）；(d) 知识图谱（用实体关系扩展）；(e) 多查询生成（生成多个改写、各自检索、融合）。与’多查询’的关系——(a) 单查询扩展（把扩展词并入原查询）；(b) 多查询（生成 m 个改写，各自检索，用 RRF 融合）；后者更鲁棒（某个改写偏了不至于全错）。评估——(a) 召回提升（扩展的收益）；(b) 精度损失（引入的噪声）；(c) 端到端 NDCG；(d) 漂移率（扩展后 top-k 与原 top-k 的重叠度——重叠过低说明漂移）。实践建议——(a) 多轮对话必做上下文补全（否则检索质量差）；(b) 短查询可扩展（配多查询 + RRF 融合）；(c) 长/精确查询不扩展；(d) 用多查询而非单查询扩展（更鲁棒）；(e) 按查询类型路由（分类后决定是否改写）；(f) 监控漂移率。度量——(a) 分查询类型的召回/NDCG；(b) 漂移率；(c) 延迟与成本。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Systematic & Methodological Formulation: Query Transformation Strategies.
(1) Classical Pseudo-Relevance Feedback (PRF / RM3):
Assumes top-$k$ retrieved documents $D_k = {d_1, dots, d_k}$ from an initial BM25 search are pseudo-relevant. It estimates a feedback language model $P(w mid theta_F)$:
$$P(w mid theta_F) = sum_{d in D_k} P(w mid d) P(d mid q)$$
The expanded query model is a linear interpolation between original query model $theta_Q$ and feedback model $theta_F$:
$$P(w mid tilde{theta}_Q) = (1 – alpha) P(w mid theta_Q) + alpha P(w mid theta_F)$$
Top expanded terms with highest probability are appended to the original query.
(2) Hypothetical Document Embeddings (HyDE, Gao et al., 2022):
For complex queries where direct question-document similarity is weak:
– Step 1: Instruct an LLM to generate a zero-shot hypothetical answer $d_{text{hypo}} sim P_{text{LLM}}(d mid q)$. (Even if factually incorrect, its language and topical structure mirror real target documents).
– Step 2: Encode the hypothetical document: $v_{text{hypo}} = E_D(d_{text{hypo}})$.
– Step 3: Retrieve real documents via vector similarity $s(d_{text{hypo}}, d_{text{real}}) = v_{text{hypo}}^T E_D(d_{text{real}})$.
(3) Multi-Turn Conversational Coreference & Disambiguation:
In conversational search, user queries contain elliptical references: $q_t = text{‘how much does it cost?’}$. An LLM rewriter conditions on session history $H_{<t}$:
$$q_{text{rewritten}} = text{LLM}(q_t, H_{<t}) implies text{'how much does Tesla Model Y cost in 2024?'}$$
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
深度剖析与工程权衡：① ‘查询漂移’是扩展的核心风险——PRF 尤其明显；面试中能指出这一点是深度理解的标志。② ‘不是所有查询都该改写’——长/精确查询扩展会引入噪声；故需’按查询类型路由’。③ ‘多查询 + RRF 融合’比’单查询扩展’更鲁棒——因为某个改写偏了不至于全错。④ ‘多轮对话的上下文补全’是必需的——否则’它是什么’这类查询无法检索；这是对话式检索的基础。⑤ ‘多跳分解’与’迭代检索’的关系——多跳问题需’边推理边检索’（见 Self-RAG 与多跳题）；不是一次改写能解决的。⑥ 面试要点——被问’查询改写有什么用’，应给出’改写（补全/纠错/分解）与扩展（同义词/PRF）+ 收益（召回）+ 风险（漂移/噪声/成本）‘与’按查询类型路由 + 多查询融合 + 监控漂移率‘；能指出’不是所有查询都该改写’是深度理解的标志。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
In-Depth Analysis & Engineering Trade-offs: ① The query drift risk—if the initial retrieval or LLM hallucination is topically flawed, PRF or HyDE severely magnifies the error, pulling irrelevant documents into the recall set; selective rewriting based on query ambiguity confidence is critical. ② Latency overhead of LLM rewriting—calling a 7B LLM to rewrite or expand queries introduces 100–300ms latency, making it unsuitable for 50ms search SLAs; production search deploys distilled 100M-parameter sequence-to-sequence models or cached rewrite dictionaries for online paths. ③ Exact keyword preservation—rewriting must never alter or drop critical technical tokens (product SKUs, error numbers, exact names); production rewriters lock recognized named entities. ④ Multi-query parallel search—generating 3 diverse query reformulations and executing parallel searches followed by RRF fusion significantly improves recall on ambiguous queries at the cost of 3x search compute. ⑤ HyDE vs. query-to-query dense retrieval—HyDE bridges the asymmetric question-passage gap by converting search into a symmetric document-to-document match, excelling at complex analytical queries. ⑥ Interview takeaway—distinguish query rewriting (spelling, coreference, intent decomposition) from query expansion (synonym injection, PRF, HyDE), explain the query drift failure mode, and discuss latency mitigation strategies.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 对所有查询都做扩展（长/精确查询受损）
- ⚠️ 用 PRF 不监控漂移（可能越走越偏）
English Pitfalls:
– Applying blind query expansion to specific, long-tail technical queries, introducing semantic drift and degrading precision.
– Deploying heavy generative LLM rewriters directly on synchronous latency-critical search pipelines without aggressive caching or distillation.
– Permitting query rewriting models to alter or discard exact entity identifiers (e.g., error codes, model numbers).
六、高频深度面试追问与预测 (Follow-Up Questions)
- 改写与扩展的区别？
- Why does HyDE succeed in retrieving relevant documents even when the generated hypothetical passage contains factual hallucinations?
- 什么时候不该改写？
- How can a search system detect query ambiguity to decide whether to trigger query expansion dynamically?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
双路召回融合策略：倒数排名融合 (RRF) 与加权线性分数归一化(Hybrid Retrieval & Reciprocal Rank Fusion (RRF)) - 🗺️ 知识图谱模块：
AI 应用与 Agent 拓扑导图
🔬 算法科学家与机器学习深度考察全量题库 (Science Depth)
本题收录于 TalentMe 算法科学家深度考察真题库 (Science Depth)。全库共 856 道硬核考点，深度覆盖数学统计、经典ML、深度学习、Transformer、大语言模型、多模态、推荐系统与 MLOps。支持 Jev 面经智能匹配、一键离线单文件 HTML 手册导出并直连 Obsidian 本地记忆。