_{🧠 图解记忆：先找正确证据在哪一层丢失，再修改对应环节；点击图片可查看原图。}

💡 答案要点

问题分析:

召回率低 = 相关文档没检索到

可能原因:
1. 分块策略不合理
2. Embedding模型不适配
3. 只用单一检索方式
4. Query表达不清晰

诊断步骤:

Step 1: 建立评估数据集

# 构建100条标准问答对
test_dataset = [
    {
        "question": "员工请假超过3天需要什么审批?",
        "ground_truth_docs": ["doc_123", "doc_456"],  # 应该检索到的文档ID
        "ideal_answer": "超过3天需要部门经理+HR审批"
    },
    # ... 100条
]

# 计算Recall@K
def evaluate_recall(retriever, dataset, k=5):
    total_recall = 0
    for item in dataset:
        retrieved = retriever.search(item["question"], k=k)
        retrieved_ids = [doc.id for doc in retrieved]

        # 计算召回了多少ground_truth
        hits = len(set(retrieved_ids) & set(item["ground_truth_docs"]))
        recall = hits / len(item["ground_truth_docs"])
        total_recall += recall

    return total_recall / len(dataset)

# 基线: Recall@5 = 0.68

Step 2: 逐项优化

(1) 优化分块 (+11% Recall)

# Before: 固定长度
splitter_old = RecursiveCharacterTextSplitter(chunk_size=512)

# After: 语义分块 + 元数据
from langchain.text_splitter import MarkdownTextSplitter

splitter_new = MarkdownTextSplitter(
    chunk_size=800,
    chunk_overlap=200,
    # 保留标题层级
    add_start_index=True
)

# 添加元数据
for chunk in chunks:
    chunk.metadata = {
        "source": doc.source,
        "title": extract_title(chunk),
        "section": extract_section(chunk),
        "page": chunk.page
    }

# Recall: 0.68 → 0.79 (+11%)

(2) 混合检索 (+8% Recall)

# Before: 只用向量检索
vector_results = vector_db.search(query)

# After: BM25 + Vector + RRF
def hybrid_retrieval(query, k=20):
    # BM25擅长关键词
    bm25_results = bm25.search(query, k=k)

    # Vector擅长语义
    vector_results = vector_db.search(
        embed(query), k=k
    )

    # RRF融合
    return rrf_fusion([bm25_results, vector_results], k=60)

# Recall: 0.79 → 0.87 (+8%)

(3) Query改写 (+5% Recall)

展开 Python 代码示例（33 行）

# Before: 直接用用户原始问题
query = "请假流程"  # 太简短,语义不明

# After: LLM扩展query
def expand_query(query):
    prompt = f"""
    原始问题: {query}

    请生成3个相关的扩展查询,帮助检索:
    1. 同义词替换
    2. 补充上下文
    3. 细化问题

    示例:
    原始: "请假流程"
    扩展:
    1. "员工请假审批流程"
    2. "如何申请休假"
    3. "请假需要哪些步骤"
    """

    expanded = llm.generate(prompt)
    return [query] + expanded  # 返回原始+扩展

# 用多个query检索,合并结果
all_results = []
for q in expand_query(query):
    all_results.extend(hybrid_retrieval(q, k=10))

# 去重+排序
final = deduplicate_and_rerank(all_results)

# Recall: 0.87 → 0.92 (+5%)

(4) Rerank精排 (+3% 保持top-5质量)

from sentence_transformers import CrossEncoder

reranker = CrossEncoder('BAAI/bge-reranker-large')

# 先召回20个候选
candidates = hybrid_retrieval(query, k=20)

# Rerank精选top-5
scores = reranker.predict([
    [query, doc.content] for doc in candidates
])

top5 = sorted(
    zip(candidates, scores),
    key=lambda x: x[1],
    reverse=True
)[:5]

# Precision@5: 0.73 → 0.89 (+16%)

最终效果:

优化路径:
基线 (0.68)
→ 语义分块 (0.79, +11%)
→ 混合检索 (0.87, +8%)
→ Query扩展 (0.92, +5%)
→ Rerank (精确率+16%)

总提升: Recall@5 从 68% → 92% (+24个百分点)

成本分析:

优化成本:
- 语义分块: 0 (只是代码改动)
- 混合检索: +100ms延迟, $0成本
- Query扩展: +$0.002/次 (LLM调用)
- Rerank: +200ms延迟, $0成本

投入产出比: 极高

面试话术:

"召回率低我先建立评估数据集量化问题,然后逐项优化:语义分块+11%,混合检索+8%,Query扩展+5%,Rerank保证精确率。最终Recall@5从68%提升到92%,代价只是+300ms延迟和每次$0.002的Query扩展成本。"

📚 参考：Anthropic：Contextual Retrieval（召回率优化实践）

---