Query Understanding——HyDE、Multi-Query、Query Expansion实战

日期: 2026-09-18 方向: AI系统工程 / RAG 阶段: Phase 3 - RAG高级模式 (Day 135-148) 标签: #QueryRewriting #HyDE #MultiQuery #QueryExpansion #LLM

---

今日目标

核心结论预告：Multi-Query在金融query上效果最稳定 (+5% Recall)，HyDE对长上下文问答最强 (+8%)，Term Expansion几乎免费但只对含缩写的query有效 (+15% on those)。生产组合：先做term expansion (cheap)，再multi-query/HyDE (有损latency但高质量)。

---

一、核心概念

1.1 用户query的"病态"分布

实际生产query长这样：

"AAPL Q4?"                     ← 极短，无上下文
"impact of fed rate hike?"     ← 含糊，对哪个公司？
"warranty 2024"                ← 关键词碎片
"what's their MOIC?"           ← 内行术语，但缺主语
"comparison of EPS and FCF"    ← 多概念但缺时段

直接embed这些query：

• "AAPL Q4?" → embedding向量靠近"AAPL"和"Q4"概念，召回杂乱
• 用户实际想问"Apple's Q4 2024 financial performance"

1.2 三大query rewrite技术

1.3 HyDE算法

Precise Zero-Shot Dense Retrieval without Relevance Labels (Gao et al., 2022)

原始 query: "AAPL Q4 services revenue?"
       ↓ LLM生成假设答案
假设答案: "Apple's Q4 2024 services revenue was approximately
         $24.5 billion, growing 11% year-over-year, driven by
         App Store, iCloud, and Apple Music subscriptions..."
       ↓ embed假设答案 (而非原query)
hypothesis_embedding
       ↓ 检索
Top-K chunks (更相关，因为假设答案embedding在向量空间中已"位于"答案chunks附近)

关键洞察：embedding model训练目标是让"语义相似"的文本靠近，"长答案"和"该答案对应的源文档"在向量空间靠近的程度比"短query"和"源文档"更高。所以即使LLM编造的假设答案有误，其embedding依然帮助找到正确chunks。

1.4 Multi-Query

原始 query: "How does the Fed affect tech stocks?"
       ↓ LLM生成5个变体
Q1: "Impact of Federal Reserve interest rate decisions on technology sector"
Q2: "Why are growth stocks sensitive to Fed monetary policy?"
Q3: "Discount rate effect on tech company valuations"
Q4: "Fed rate hikes and Nasdaq performance correlation"
Q5: "Technology stocks reaction to FOMC meetings"
       ↓ 每个query独立检索 top-10
       ↓ 合并 + 去重 + RRF
Final top-K (覆盖更全面的相关chunks)

1.5 Query Expansion (Term-based)

金融领域的缩写极多：

Dictionary lookup:
  "AAPL"   → "AAPL", "Apple Inc.", "Apple"
  "FCF"    → "FCF", "free cash flow"
  "MOIC"   → "MOIC", "multiple on invested capital"
  "Q4 2024" → "Q4 2024", "fourth quarter 2024", "fiscal Q4 2024"
  "10-K"   → "10-K", "annual report", "Form 10-K"

可以是 静态字典 或 LLM动态扩展。

1.6 Step-Back Prompting (Google DeepMind)

原始 query: "Was Apple's Q3 2024 services growth slower than its Q2 2024 services growth?"
       ↓ Step-back 抽象
背景query: "Apple's quarterly services revenue trends in fiscal 2024"
       ↓ 用background query检索
拿到全季度数据 → LLM对比回答原问题

适合"事实对比"类query。

---

二、完整实现：query_rw.py

"""
query_rw.py — 3 query rewriting techniques + ensemble
依赖：
  pip install anthropic openai chromadb numpy
"""
import os
import time
import json
from typing import List, Dict, Tuple
from anthropic import Anthropic
from openai import OpenAI

anthropic_client = Anthropic()
openai_client = OpenAI()


# ============================================================
# 1. HyDE
# ============================================================
HYDE_PROMPT = """You are a financial analyst. Given the following question,
write a concise but realistic answer of 3-5 sentences as if you were answering
based on a real document. Use specific terminology and likely figures.
DO NOT preface or explain — just write the hypothetical answer text.

Question: {query}

Hypothetical Answer:"""


def hyde_rewrite(query: str) -> str:
    resp = anthropic_client.messages.create(
        model="claude-haiku-4-5-20250929",  # 用快速便宜model
        max_tokens=200,
        messages=[{"role": "user", "content": HYDE_PROMPT.format(query=query)}],
    )
    return resp.content[0].text.strip()


# ============================================================
# 2. Multi-Query
# ============================================================
MULTI_QUERY_PROMPT = """Generate 5 alternative ways to phrase the following
financial question, capturing different angles and using diverse terminology.
Output ONLY a JSON array of 5 strings, no other text.

Original question: {query}

Output:"""


def multi_query_rewrite(query: str, n: int = 5) -> List[str]:
    resp = anthropic_client.messages.create(
        model="claude-haiku-4-5-20250929",
        max_tokens=400,
        messages=[{"role": "user", "content": MULTI_QUERY_PROMPT.format(query=query)}],
    )
    text = resp.content[0].text.strip()
    try:
        variants = json.loads(text[text.index("["):text.rindex("]") + 1])
    except Exception:
        variants = [query]
    return [query] + variants[:n]   # 包含原query


# ============================================================
# 3. Term Expansion (静态字典 + LLM)
# ============================================================
FINANCE_DICT = {
    "AAPL": ["Apple Inc.", "Apple"],
    "MSFT": ["Microsoft Corporation", "Microsoft"],
    "GOOG": ["Alphabet", "Google"],
    "NVDA": ["NVIDIA Corporation", "NVIDIA"],
    "TSLA": ["Tesla Inc.", "Tesla"],
    "JPM": ["JPMorgan Chase", "JPMorgan"],
    "FCF": ["free cash flow"],
    "EPS": ["earnings per share"],
    "MOIC": ["multiple on invested capital"],
    "IRR": ["internal rate of return"],
    "TTM": ["trailing twelve months"],
    "YoY": ["year-over-year", "year over year"],
    "QoQ": ["quarter-over-quarter"],
    "10-K": ["annual report", "Form 10-K"],
    "10-Q": ["quarterly report", "Form 10-Q"],
    "MD&A": ["management discussion and analysis"],
    "GAAP": ["generally accepted accounting principles"],
    "Fed": ["Federal Reserve"],
    "FOMC": ["Federal Open Market Committee"],
    "VIX": ["volatility index"],
    "EBITDA": ["earnings before interest, taxes, depreciation, and amortization"],
}


def term_expansion(query: str, use_llm: bool = False) -> str:
    """字典展开 + 可选LLM补充"""
    expanded = query
    for term, expansions in FINANCE_DICT.items():
        # 边界匹配
        import re
        pattern = r"\b" + re.escape(term) + r"\b"
        if re.search(pattern, expanded, re.IGNORECASE):
            for exp in expansions:
                expanded += f" ({exp})"

    if use_llm:
        # LLM补充未覆盖的术语
        prompt = f"""For the following financial question, identify any acronyms
or technical terms that might be ambiguous and add their full forms in parens.
Keep the original wording.

Question: {expanded}

Expanded:"""
        resp = anthropic_client.messages.create(
            model="claude-haiku-4-5-20250929",
            max_tokens=200,
            messages=[{"role": "user", "content": prompt}],
        )
        expanded = resp.content[0].text.strip()
    return expanded


# ============================================================
# 4. Step-Back
# ============================================================
STEP_BACK_PROMPT = """You are an expert at converting specific financial
questions to broader background questions that retrieve more contextual
information. Given the original question, output a single broader question.

Original: {query}

Broader question:"""


def step_back(query: str) -> str:
    resp = anthropic_client.messages.create(
        model="claude-haiku-4-5-20250929",
        max_tokens=100,
        messages=[{"role": "user", "content": STEP_BACK_PROMPT.format(query=query)}],
    )
    return resp.content[0].text.strip()


# ============================================================
# 5. Retrieval with rewritten query
# ============================================================
import chromadb

def get_collection():
    client = chromadb.PersistentClient(path="./chroma_db")
    return client.get_collection("financial_docs_v1")


def embed(text: str) -> List[float]:
    return openai_client.embeddings.create(
        model="text-embedding-3-large", input=[text]
    ).data[0].embedding


def retrieve(query: str, top_k: int = 10) -> List[Dict]:
    coll = get_collection()
    res = coll.query(query_embeddings=[embed(query)], n_results=top_k)
    return [
        {"id": res["ids"][0][i],
         "text": res["documents"][0][i],
         "distance": res["distances"][0][i]}
        for i in range(len(res["ids"][0]))
    ]


# ============================================================
# 6. Query Rewriting Strategies
# ============================================================
def retrieve_baseline(query: str, top_k: int = 10) -> List[Dict]:
    return retrieve(query, top_k=top_k)


def retrieve_hyde(query: str, top_k: int = 10) -> List[Dict]:
    hyp = hyde_rewrite(query)
    return retrieve(hyp, top_k=top_k)


def retrieve_term_expand(query: str, top_k: int = 10) -> List[Dict]:
    expanded = term_expansion(query, use_llm=False)
    return retrieve(expanded, top_k=top_k)


def retrieve_multi_query(query: str, top_k: int = 10) -> List[Dict]:
    variants = multi_query_rewrite(query, n=5)
    all_results = {}
    for i, v in enumerate(variants):
        results = retrieve(v, top_k=top_k)
        for rank, r in enumerate(results):
            cid = r["id"]
            # RRF
            all_results[cid] = all_results.get(cid, {"text": r["text"], "rrf": 0})
            all_results[cid]["rrf"] += 1 / (60 + rank)

    sorted_ids = sorted(all_results.items(), key=lambda x: -x[1]["rrf"])
    return [{"id": k, "text": v["text"], "rrf_score": v["rrf"]}
            for k, v in sorted_ids[:top_k]]


def retrieve_ensemble(query: str, top_k: int = 10) -> List[Dict]:
    """组合：term expand → 在expanded query上跑multi-query + HyDE → RRF合并"""
    expanded = term_expansion(query, use_llm=False)

    multi_results = retrieve_multi_query(expanded, top_k=20)
    hyde_results = retrieve_hyde(expanded, top_k=20)

    rrf = {}
    for rank, r in enumerate(multi_results):
        rrf[r["id"]] = rrf.get(r["id"], {"text": r["text"], "score": 0})
        rrf[r["id"]]["score"] += 1 / (60 + rank)
    for rank, r in enumerate(hyde_results):
        rrf[r["id"]] = rrf.get(r["id"], {"text": r["text"], "score": 0})
        rrf[r["id"]]["score"] += 1 / (60 + rank)

    sorted_ids = sorted(rrf.items(), key=lambda x: -x[1]["score"])
    return [{"id": k, "text": v["text"], "score": v["score"]}
            for k, v in sorted_ids[:top_k]]


# ============================================================
# 7. 评估
# ============================================================
import numpy as np

def evaluate(retrieve_fn, queries: List[Dict]) -> Dict:
    recall_5, mrr_list, latencies = [], [], []
    for q in queries:
        t0 = time.time()
        results = retrieve_fn(q["query"], top_k=10)
        latencies.append((time.time() - t0) * 1000)

        top_ids = [r["id"] for r in results]
        gt = set(q["ground_truth_ids"])
        recall_5.append(len(gt & set(top_ids[:5])) / len(gt))
        rr = next((1/(rank+1) for rank, cid in enumerate(top_ids) if cid in gt), 0)
        mrr_list.append(rr)

    return {
        "recall@5": float(np.mean(recall_5)),
        "MRR": float(np.mean(mrr_list)),
        "p50_ms": float(np.percentile(latencies, 50)),
    }


def main():
    with open("benchmark_dataset.json") as f:
        bench = json.load(f)

    methods = {
        "baseline": retrieve_baseline,
        "term_expand": retrieve_term_expand,
        "hyde": retrieve_hyde,
        "multi_query": retrieve_multi_query,
        "ensemble": retrieve_ensemble,
    }
    for name, fn in methods.items():
        m = evaluate(fn, bench["queries"])
        print(f"{name:15s} | Recall@5: {m['recall@5']:.3f} | "
              f"MRR: {m['MRR']:.3f} | p50: {m['p50_ms']:.0f}ms")


if __name__ == "__main__":
    main()

---

三、实测结果

3.1 在50对金融query上的对比

观察：

• term_expand几乎免费但只 +2% Recall
• HyDE显著提升 +6% Recall
• Multi-Query最强 +7% Recall，但5x latency
• Ensemble +10% Recall，但1.5s latency太高

3.2 按query类型分层

关键洞察：

• 缩写/代码query → term_expand最强
• 复杂长query → HyDE最强
• 短模糊query → multi_query最强
• 简单query不做rewrite也行

3.3 真实例子对比

Q: "AAPL FCF 2024?"

baseline 召回:
  apple_10k_p44_c1: revenue trends (general)
  apple_10k_p43_c2: products gross margin
  apple_10k_p15_c3: AAPL forward guidance
  → 找到FCF的rank ≥ 8

+ term_expand "AAPL (Apple Inc.) FCF (free cash flow) 2024":
  apple_10k_p47_c1: cash flow statement ✓
  apple_10k_p46_c2: operating activities ✓
  → top-2即命中

---

四、金融领域应用

4.1 案例：Earnings Call问答

Q: "What did the CEO say about AI investments?"

HyDE生成的假设答案：
"During the earnings call, the CEO emphasized continued investment in
artificial intelligence capabilities, including ongoing development of
proprietary models, infrastructure expansion, and partnerships with
leading AI research institutions. The company allocated approximately
$X billion in AI-related capex..."

→ embedding该假设答案，找到实际call transcript里的相关片段。

→ HyDE特别适合 会有具体技术性回答 的query。

4.2 监管法规RAG的query rewrite

Q: "我的客户能不能因为KYC不全就被拒绝开户？" (中文)

term_expand:
  "KYC (Know Your Customer)"
  "不全 (incomplete)"
  "开户 (account opening)"

translate + multi_query:
  Q1: "Can KYC requirements lead to account opening rejection?"
  Q2: "Customer due diligence failure consequences for new accounts"
  Q3: "AML KYC documentation requirements for retail clients"

跨语言 + 法规RAG，ensemble方案最合适。

4.3 投研问答的Step-Back

Q: "Compare Apple's Q3 2024 to Q2 2024 services growth"

Step-Back: "Apple's quarterly services revenue trends in fiscal 2024"
→ 检索拿到全4个季度数据
→ LLM在context里对比Q3 vs Q2

→ 直接query只能命中Q3或Q2的某一段，step-back拿到全数据更可靠。

---

五、生产经验

5.1 8个query rewrite的坑

5.2 异步Multi-Query

import asyncio

async def retrieve_async(query):
    # ...
    pass

async def multi_query_async(query: str, top_k: int = 10):
    variants = multi_query_rewrite(query)
    tasks = [retrieve_async(v) for v in variants]
    results = await asyncio.gather(*tasks)
    # RRF融合
    return rrf_merge(results, top_k)

→ 5次retrieve从串行1.2s变并行250ms。

5.3 Cache rewrite结果

@lru_cache(maxsize=10000)
def cached_hyde(query: str) -> str:
    return hyde_rewrite(query)

但要注意：

• 大小写、空格规范化
• 同义query不同写法仍重复
• 用 embedding-based cache：query embed 距离 < 0.05 用cached

---

六、Cost & Latency

6.1 不同方法的成本（10K queries/day）

关键：异步可以让multi-query延迟代价从1.2s降到30ms！

6.2 何时用哪个的决策矩阵

---

七、关键速查表

7.1 Query Rewrite方法对比

7.2 Production组合策略

[Query] ────────────────────────┐
   │                            │
   ▼                            │
[Term Expand] (always-on)       │
   │                            │
   ▼                            │
[query_length > 5 words?]       │
   │                            │
   ├── yes → [HyDE]             │
   │           │                │
   │           ▼                │
   │       [Retrieve top-50]    │
   │                            │
   └── no → [Multi-Query async] │
              │                 │
              ▼                 │
          [Retrieve top-50]     │
              │                 │
              ▼                 │
          [bge-rerank top-5]    │
              │                 │
              ▼                 │
          [Generate]            │

---

八、面试题

Q1: HyDE的核心idea为什么work？LLM不是会幻觉吗？

HyDE的关键不是 答案的内容正确，而是 答案的embedding在向量空间中位置。LLM编造的假设答案，即使数字错了，但写作风格、专业术语、句法结构都接近真实答案文档。embedding model对这些"answer-like text"的encoding比对短query的encoding更接近真实答案chunk。所以哪怕"Apple Q4 revenue $200B"是错的，假设答案的embedding依然帮助找到真实"Apple Q4 revenue $94.93B"的chunk。HyDE论文显示在TREC-DL等benchmark上 +5-10% nDCG。

Q2: Multi-Query的5个变体怎么保证diversity？

三个手段：(1) prompt显式diverse：要求"different angles, diverse terminology"; (2) temperature=0.7-0.9：增加随机性; (3) chain-of-thought：先让LLM列举不同的"信息需求维度"再生成query；(4) post-hoc dedup：用embedding sim>0.95的过滤掉重复。生产上简单(1)+(2)就够，重要的是 保留原query 在variants里防止LLM rewrite过度。

Q3: 你的RAG有10K query/day，加HyDE后latency从0.3s变1s，怎么办？

三步：(1) 先看是否值得：HyDE的lift如果<3%，不要加；如果>5%，继续优化; (2) 换Haiku或更小模型做rewrite：HyDE对rewrite quality要求不高，小模型够; (3) 并行rewrite和初步retrieve：在LLM rewrite的同时跑baseline retrieve，然后合并；(4) 缓存高频rewrite结果：财经类query重复率高，缓存命中率可达30-50%。如果都不够，回退到term_expand only。

Q4: Step-Back vs HyDE，分别适合什么场景？

Step-Back：query非常 具体，且需要的答案可能要 多chunk组合。例："Was Q3 services growth slower than Q2?" → 抽象到"Q1-Q4 services growth"再对比。HyDE：query 通用 但需要 专业术语丰富的答案。例："impact of Fed rate on tech stocks" → HyDE生成包含"discount rate", "DCF model", "growth multiples"的假设答案，更接近真实研究报告。

Q5: 给你一个新的RAG项目，怎么决定要不要做query rewrite？

三步评估：(1) 看query分布：随机抽100个用户query，统计 (a) 平均长度 (b) 是否含缩写 (c) 是否含specific entities。如果 <10 words 占 50%+ 或缩写多，rewrite收益大。(2) 跑A/B测试：50 queries上baseline vs rewrite，看Recall@5和用户满意度。如果lift > 5% 才值得加。(3) 看latency budget：如果<500ms必须达成，term_expand only；可允许>1s，可考虑HyDE+multi_query ensemble。

---

九、明日预告

Day 141: Week 21复习 + RAG v2整合——把Day 135-140的所有改进（embedding选型、vector DB、hybrid search、reranking、query rewrite）整合成 rag_v2.py，端到端在benchmark上跑分。预计Recall@5从v1的0.864提升到 0.948+。同时总结RAG调优方法论的"先做什么后做什么"决策框架。