RAG检索增强优化指南

目录

• RAG检索增强优化指南
• • 一、三代RAG架构对比
  • • 1.1 架构演进时间线
    • 1.2 三代架构详细对比
    • 1.3 Naive RAG – 基础架构
    • • Python实现示例
    • 1.4 Advanced RAG – 增强架构
    • 1.5 Modular RAG – 模块化架构
    • 1.6 RAG 全链路增强路线
  • 二、检索前增强
  • • 2.1 HyDE (Hypothetical Document Embeddings)
    • 2.2 Query2Doc
    • 2.3 查询重写 (Query Rewriting)
    • 2.4 查询分解 (Query Decomposition)
    • 2.5 Take Step Back
  • 三、检索增强
  • • 3.1 核心概念
    • 3.2 混合检索 (Hybrid Search)
    • • BM25算法原理
      • 向量检索原理
      • RRF融合算法
    • 3.3 句子窗口检索 (Sentence Window Retriever)
    • 3.4 父子文档检索 (Parent Document Retriever)
    • 3.5 摘要检索 (Summary Retriever)
    • 3.6 假设性问题检索 (Hypothetical Questions Retriever)
    • 3.7 多索引融合策略
  • 四、检索后增强
  • • 4.1 重排序 (Re-ranking) 原理
    • 4.2 主流Re-ranker对比
    • 4.3 Re-ranking实现
    • 4.4 上下文压缩
  • 五、完整生产级示例
  • • 5.1 LangChain生产级实现
    • 5.2 LlamaIndex生产级实现

RAG检索增强优化指南

核心概念 RAG (Retrieval-Augmented Generation) 是一种革命性的AI架构，巧妙地将检索系统与生成式大语言模型结合。它像给大模型配备了一个"外挂知识库"，让AI能够：

• 从外部知识库实时获取信息
• 基于事实生成答案，大幅减少幻觉
• 知识库可随时更新，无需重新训练模型
• 提供答案来源，增强可解释性

一、三代RAG架构对比

RAG技术经历了三代演进，每一代都解决了前一代的痛点。

1.1 架构演进时间线

#mermaid-svg-cEUNC1ekp3T38YEU{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-cEUNC1ekp3T38YEU .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-cEUNC1ekp3T38YEU .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-cEUNC1ekp3T38YEU .error-icon{fill:#552222;}#mermaid-svg-cEUNC1ekp3T38YEU .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-cEUNC1ekp3T38YEU .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-cEUNC1ekp3T38YEU .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-cEUNC1ekp3T38YEU .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-cEUNC1ekp3T38YEU .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-cEUNC1ekp3T38YEU .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-cEUNC1ekp3T38YEU .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-cEUNC1ekp3T38YEU .marker{fill:#333333;stroke:#333333;}#mermaid-svg-cEUNC1ekp3T38YEU .marker.cross{stroke:#333333;}#mermaid-svg-cEUNC1ekp3T38YEU svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-cEUNC1ekp3T38YEU p{margin:0;}#mermaid-svg-cEUNC1ekp3T38YEU .edge{stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .section–1 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section–1 path,#mermaid-svg-cEUNC1ekp3T38YEU .section–1 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section–1 path{fill:hsl(240, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section–1 text{fill:#ffffff;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon–1{font-size:40px;color:#ffffff;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge–1{stroke:hsl(240, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth–1{stroke-width:17;}#mermaid-svg-cEUNC1ekp3T38YEU .section–1 line{stroke:hsl(60, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:#ffffff;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-0 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-0 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-0 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-0 path{fill:hsl(60, 100%, 73.5294117647%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-0 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-0{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-0{stroke:hsl(60, 100%, 73.5294117647%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-0{stroke-width:14;}#mermaid-svg-cEUNC1ekp3T38YEU .section-0 line{stroke:hsl(240, 100%, 83.5294117647%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-1 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-1 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-1 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-1 path{fill:hsl(80, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-1 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-1{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-1{stroke:hsl(80, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-1{stroke-width:11;}#mermaid-svg-cEUNC1ekp3T38YEU .section-1 line{stroke:hsl(260, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-2 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-2 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-2 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-2 path{fill:hsl(270, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-2 text{fill:#ffffff;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-2{font-size:40px;color:#ffffff;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-2{stroke:hsl(270, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-2{stroke-width:8;}#mermaid-svg-cEUNC1ekp3T38YEU .section-2 line{stroke:hsl(90, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:#ffffff;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-3 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-3 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-3 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-3 path{fill:hsl(300, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-3 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-3{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-3{stroke:hsl(300, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-3{stroke-width:5;}#mermaid-svg-cEUNC1ekp3T38YEU .section-3 line{stroke:hsl(120, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-4 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-4 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-4 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-4 path{fill:hsl(330, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-4 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-4{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-4{stroke:hsl(330, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-4{stroke-width:2;}#mermaid-svg-cEUNC1ekp3T38YEU .section-4 line{stroke:hsl(150, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-5 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-5 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-5 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-5 path{fill:hsl(0, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-5 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-5{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-5{stroke:hsl(0, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-5{stroke-width:-1;}#mermaid-svg-cEUNC1ekp3T38YEU .section-5 line{stroke:hsl(180, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-6 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-6 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-6 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-6 path{fill:hsl(30, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-6 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-6{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-6{stroke:hsl(30, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-6{stroke-width:-4;}#mermaid-svg-cEUNC1ekp3T38YEU .section-6 line{stroke:hsl(210, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-7 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-7 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-7 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-7 path{fill:hsl(90, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-7 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-7{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-7{stroke:hsl(90, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-7{stroke-width:-7;}#mermaid-svg-cEUNC1ekp3T38YEU .section-7 line{stroke:hsl(270, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-8 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-8 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-8 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-8 path{fill:hsl(150, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-8 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-8{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-8{stroke:hsl(150, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-8{stroke-width:-10;}#mermaid-svg-cEUNC1ekp3T38YEU .section-8 line{stroke:hsl(330, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-9 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-9 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-9 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-9 path{fill:hsl(180, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-9 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-9{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-9{stroke:hsl(180, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-9{stroke-width:-13;}#mermaid-svg-cEUNC1ekp3T38YEU .section-9 line{stroke:hsl(0, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-10 rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-10 path,#mermaid-svg-cEUNC1ekp3T38YEU .section-10 circle,#mermaid-svg-cEUNC1ekp3T38YEU .section-10 path{fill:hsl(210, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-10 text{fill:black;}#mermaid-svg-cEUNC1ekp3T38YEU .node-icon-10{font-size:40px;color:black;}#mermaid-svg-cEUNC1ekp3T38YEU .section-edge-10{stroke:hsl(210, 100%, 76.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .edge-depth-10{stroke-width:-16;}#mermaid-svg-cEUNC1ekp3T38YEU .section-10 line{stroke:hsl(30, 100%, 86.2745098039%);stroke-width:3;}#mermaid-svg-cEUNC1ekp3T38YEU .lineWrapper line{stroke:black;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled,#mermaid-svg-cEUNC1ekp3T38YEU .disabled circle,#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:lightgray;}#mermaid-svg-cEUNC1ekp3T38YEU .disabled text{fill:#efefef;}#mermaid-svg-cEUNC1ekp3T38YEU .section-root rect,#mermaid-svg-cEUNC1ekp3T38YEU .section-root path,#mermaid-svg-cEUNC1ekp3T38YEU .section-root circle{fill:hsl(240, 100%, 46.2745098039%);}#mermaid-svg-cEUNC1ekp3T38YEU .section-root text{fill:#ffffff;}#mermaid-svg-cEUNC1ekp3T38YEU .icon-container{height:100%;display:flex;justify-content:center;align-items:center;}#mermaid-svg-cEUNC1ekp3T38YEU .edge{fill:none;}#mermaid-svg-cEUNC1ekp3T38YEU .eventWrapper{filter:brightness(120%);}#mermaid-svg-cEUNC1ekp3T38YEU :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

2022

Naive RAG (基础版)

简单检索+生成

2023

Advanced RAG

(增强版)

查询增强+混合检索+重排序

2024

Modular RAG

(模块化版)

动态路由+自我评估

RAG 架构演进史

1.2 三代架构详细对比

1.3 Naive RAG – 基础架构

工作原理：最简单的RAG实现，就像"查字典-翻译"的过程：

#mermaid-svg-woA7l6vtMOin57Cp{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-woA7l6vtMOin57Cp .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-woA7l6vtMOin57Cp .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-woA7l6vtMOin57Cp .error-icon{fill:#552222;}#mermaid-svg-woA7l6vtMOin57Cp .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-woA7l6vtMOin57Cp .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-woA7l6vtMOin57Cp .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-woA7l6vtMOin57Cp .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-woA7l6vtMOin57Cp .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-woA7l6vtMOin57Cp .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-woA7l6vtMOin57Cp .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-woA7l6vtMOin57Cp .marker{fill:#333333;stroke:#333333;}#mermaid-svg-woA7l6vtMOin57Cp .marker.cross{stroke:#333333;}#mermaid-svg-woA7l6vtMOin57Cp svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-woA7l6vtMOin57Cp p{margin:0;}#mermaid-svg-woA7l6vtMOin57Cp .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-woA7l6vtMOin57Cp .cluster-label text{fill:#333;}#mermaid-svg-woA7l6vtMOin57Cp .cluster-label span{color:#333;}#mermaid-svg-woA7l6vtMOin57Cp .cluster-label span p{background-color:transparent;}#mermaid-svg-woA7l6vtMOin57Cp .label text,#mermaid-svg-woA7l6vtMOin57Cp span{fill:#333;color:#333;}#mermaid-svg-woA7l6vtMOin57Cp .node rect,#mermaid-svg-woA7l6vtMOin57Cp .node circle,#mermaid-svg-woA7l6vtMOin57Cp .node ellipse,#mermaid-svg-woA7l6vtMOin57Cp .node polygon,#mermaid-svg-woA7l6vtMOin57Cp .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-woA7l6vtMOin57Cp .rough-node .label text,#mermaid-svg-woA7l6vtMOin57Cp .node .label text,#mermaid-svg-woA7l6vtMOin57Cp .image-shape .label,#mermaid-svg-woA7l6vtMOin57Cp .icon-shape .label{text-anchor:middle;}#mermaid-svg-woA7l6vtMOin57Cp .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-woA7l6vtMOin57Cp .rough-node .label,#mermaid-svg-woA7l6vtMOin57Cp .node .label,#mermaid-svg-woA7l6vtMOin57Cp .image-shape .label,#mermaid-svg-woA7l6vtMOin57Cp .icon-shape .label{text-align:center;}#mermaid-svg-woA7l6vtMOin57Cp .node.clickable{cursor:pointer;}#mermaid-svg-woA7l6vtMOin57Cp .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-woA7l6vtMOin57Cp .arrowheadPath{fill:#333333;}#mermaid-svg-woA7l6vtMOin57Cp .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-woA7l6vtMOin57Cp .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-woA7l6vtMOin57Cp .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-woA7l6vtMOin57Cp .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-woA7l6vtMOin57Cp .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-woA7l6vtMOin57Cp .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-woA7l6vtMOin57Cp .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-woA7l6vtMOin57Cp .cluster text{fill:#333;}#mermaid-svg-woA7l6vtMOin57Cp .cluster span{color:#333;}#mermaid-svg-woA7l6vtMOin57Cp div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-woA7l6vtMOin57Cp .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-woA7l6vtMOin57Cp rect.text{fill:none;stroke-width:0;}#mermaid-svg-woA7l6vtMOin57Cp .icon-shape,#mermaid-svg-woA7l6vtMOin57Cp .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-woA7l6vtMOin57Cp .icon-shape p,#mermaid-svg-woA7l6vtMOin57Cp .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-woA7l6vtMOin57Cp .icon-shape rect,#mermaid-svg-woA7l6vtMOin57Cp .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-woA7l6vtMOin57Cp .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-woA7l6vtMOin57Cp .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-woA7l6vtMOin57Cp :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

用户问题

向量检索

相关文档 Top-K

构建Prompt

LLM生成答案

返回结果

专业术语解释：

• 向量检索：将文本转换为高维向量，通过计算向量相似度找到相关文档
• Top-K：检索最相关的前K个文档，通常K取3-10
• Prompt Engineering：精心设计输入给LLM的提示词，以获得更好的输出

Python实现示例

from langchain.embeddings import OpenAIEmbeddings
from langchain.vectorstores import Chroma
from langchain.chat_models import ChatOpenAI
from langchain.chains import RetrievalQA

# 步骤1: 初始化嵌入模型 – 将文本转换为向量
embeddings = OpenAIEmbeddings()

# 步骤2: 创建向量数据库 – 存储和检索文档向量
vectorstore = Chroma(
collection_name="knowledge_base",
embedding_function=embeddings
)

# 添加知识文档
documents = [
"RAG是Retrieval-Augmented Generation的缩写…",
"向量数据库是专门存储和检索向量的数据库…",
"Embedding模型将文本转换为高维向量…"
]
vectorstore.add_documents(documents)

# 步骤3: 创建RAG链
retriever = vectorstore.as_retriever(search_kwargs={"k": 3})
qa_chain = RetrievalQA.from_chain_type(
llm=ChatOpenAI(model="gpt-4", temperature=0),
retriever=retriever,
return_source_documents=True
)

# 执行查询
result = qa_chain({"query": "什么是RAG技术？"})

1.4 Advanced RAG – 增强架构

核心痛点：Naive RAG的三大问题

• 语义鸿沟：用户问题和文档用词不同，检索失败
• 精度不足：向量相似度不等于问题相关性
• 单一维度：只用语义检索，错过关键词匹配

Advanced RAG解决方案：

#mermaid-svg-0xn5cTjlm7cR6ChE{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-0xn5cTjlm7cR6ChE .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-0xn5cTjlm7cR6ChE .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-0xn5cTjlm7cR6ChE .error-icon{fill:#552222;}#mermaid-svg-0xn5cTjlm7cR6ChE .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-0xn5cTjlm7cR6ChE .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-0xn5cTjlm7cR6ChE .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-0xn5cTjlm7cR6ChE .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-0xn5cTjlm7cR6ChE .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-0xn5cTjlm7cR6ChE .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-0xn5cTjlm7cR6ChE .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-0xn5cTjlm7cR6ChE .marker{fill:#333333;stroke:#333333;}#mermaid-svg-0xn5cTjlm7cR6ChE .marker.cross{stroke:#333333;}#mermaid-svg-0xn5cTjlm7cR6ChE svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-0xn5cTjlm7cR6ChE p{margin:0;}#mermaid-svg-0xn5cTjlm7cR6ChE .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-0xn5cTjlm7cR6ChE .cluster-label text{fill:#333;}#mermaid-svg-0xn5cTjlm7cR6ChE .cluster-label span{color:#333;}#mermaid-svg-0xn5cTjlm7cR6ChE .cluster-label span p{background-color:transparent;}#mermaid-svg-0xn5cTjlm7cR6ChE .label text,#mermaid-svg-0xn5cTjlm7cR6ChE span{fill:#333;color:#333;}#mermaid-svg-0xn5cTjlm7cR6ChE .node rect,#mermaid-svg-0xn5cTjlm7cR6ChE .node circle,#mermaid-svg-0xn5cTjlm7cR6ChE .node ellipse,#mermaid-svg-0xn5cTjlm7cR6ChE .node polygon,#mermaid-svg-0xn5cTjlm7cR6ChE .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-0xn5cTjlm7cR6ChE .rough-node .label text,#mermaid-svg-0xn5cTjlm7cR6ChE .node .label text,#mermaid-svg-0xn5cTjlm7cR6ChE .image-shape .label,#mermaid-svg-0xn5cTjlm7cR6ChE .icon-shape .label{text-anchor:middle;}#mermaid-svg-0xn5cTjlm7cR6ChE .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-0xn5cTjlm7cR6ChE .rough-node .label,#mermaid-svg-0xn5cTjlm7cR6ChE .node .label,#mermaid-svg-0xn5cTjlm7cR6ChE .image-shape .label,#mermaid-svg-0xn5cTjlm7cR6ChE .icon-shape .label{text-align:center;}#mermaid-svg-0xn5cTjlm7cR6ChE .node.clickable{cursor:pointer;}#mermaid-svg-0xn5cTjlm7cR6ChE .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-0xn5cTjlm7cR6ChE .arrowheadPath{fill:#333333;}#mermaid-svg-0xn5cTjlm7cR6ChE .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-0xn5cTjlm7cR6ChE .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-0xn5cTjlm7cR6ChE .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-0xn5cTjlm7cR6ChE .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-0xn5cTjlm7cR6ChE .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-0xn5cTjlm7cR6ChE .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-0xn5cTjlm7cR6ChE .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-0xn5cTjlm7cR6ChE .cluster text{fill:#333;}#mermaid-svg-0xn5cTjlm7cR6ChE .cluster span{color:#333;}#mermaid-svg-0xn5cTjlm7cR6ChE div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-0xn5cTjlm7cR6ChE .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-0xn5cTjlm7cR6ChE rect.text{fill:none;stroke-width:0;}#mermaid-svg-0xn5cTjlm7cR6ChE .icon-shape,#mermaid-svg-0xn5cTjlm7cR6ChE .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-0xn5cTjlm7cR6ChE .icon-shape p,#mermaid-svg-0xn5cTjlm7cR6ChE .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-0xn5cTjlm7cR6ChE .icon-shape rect,#mermaid-svg-0xn5cTjlm7cR6ChE .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-0xn5cTjlm7cR6ChE .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-0xn5cTjlm7cR6ChE .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-0xn5cTjlm7cR6ChE :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

用户问题

查询增强

混合检索稀疏+密集

重排序 Re-rank

构建Prompt

LLM生成

专业术语解释：

• 查询增强：使用HyDE等技术改写用户问题，使其更易于检索
• 混合检索：结合稀疏检索（BM25）和密集检索（向量），兼顾关键词和语义
• 重排序：用专门模型对检索结果进行精排，提升Top结果质量

1.5 Modular RAG – 模块化架构

设计哲学：将RAG拆解为可插拔的模块，像搭积木一样自由组合

#mermaid-svg-v0udIuTaje09TKuj{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-v0udIuTaje09TKuj .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-v0udIuTaje09TKuj .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-v0udIuTaje09TKuj .error-icon{fill:#552222;}#mermaid-svg-v0udIuTaje09TKuj .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-v0udIuTaje09TKuj .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-v0udIuTaje09TKuj .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-v0udIuTaje09TKuj .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-v0udIuTaje09TKuj .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-v0udIuTaje09TKuj .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-v0udIuTaje09TKuj .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-v0udIuTaje09TKuj .marker{fill:#333333;stroke:#333333;}#mermaid-svg-v0udIuTaje09TKuj .marker.cross{stroke:#333333;}#mermaid-svg-v0udIuTaje09TKuj svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-v0udIuTaje09TKuj p{margin:0;}#mermaid-svg-v0udIuTaje09TKuj .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-v0udIuTaje09TKuj .cluster-label text{fill:#333;}#mermaid-svg-v0udIuTaje09TKuj .cluster-label span{color:#333;}#mermaid-svg-v0udIuTaje09TKuj .cluster-label span p{background-color:transparent;}#mermaid-svg-v0udIuTaje09TKuj .label text,#mermaid-svg-v0udIuTaje09TKuj span{fill:#333;color:#333;}#mermaid-svg-v0udIuTaje09TKuj .node rect,#mermaid-svg-v0udIuTaje09TKuj .node circle,#mermaid-svg-v0udIuTaje09TKuj .node ellipse,#mermaid-svg-v0udIuTaje09TKuj .node polygon,#mermaid-svg-v0udIuTaje09TKuj .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-v0udIuTaje09TKuj .rough-node .label text,#mermaid-svg-v0udIuTaje09TKuj .node .label text,#mermaid-svg-v0udIuTaje09TKuj .image-shape .label,#mermaid-svg-v0udIuTaje09TKuj .icon-shape .label{text-anchor:middle;}#mermaid-svg-v0udIuTaje09TKuj .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-v0udIuTaje09TKuj .rough-node .label,#mermaid-svg-v0udIuTaje09TKuj .node .label,#mermaid-svg-v0udIuTaje09TKuj .image-shape .label,#mermaid-svg-v0udIuTaje09TKuj .icon-shape .label{text-align:center;}#mermaid-svg-v0udIuTaje09TKuj .node.clickable{cursor:pointer;}#mermaid-svg-v0udIuTaje09TKuj .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-v0udIuTaje09TKuj .arrowheadPath{fill:#333333;}#mermaid-svg-v0udIuTaje09TKuj .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-v0udIuTaje09TKuj .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-v0udIuTaje09TKuj .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-v0udIuTaje09TKuj .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-v0udIuTaje09TKuj .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-v0udIuTaje09TKuj .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-v0udIuTaje09TKuj .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-v0udIuTaje09TKuj .cluster text{fill:#333;}#mermaid-svg-v0udIuTaje09TKuj .cluster span{color:#333;}#mermaid-svg-v0udIuTaje09TKuj div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-v0udIuTaje09TKuj .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-v0udIuTaje09TKuj rect.text{fill:none;stroke-width:0;}#mermaid-svg-v0udIuTaje09TKuj .icon-shape,#mermaid-svg-v0udIuTaje09TKuj .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-v0udIuTaje09TKuj .icon-shape p,#mermaid-svg-v0udIuTaje09TKuj .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-v0udIuTaje09TKuj .icon-shape rect,#mermaid-svg-v0udIuTaje09TKuj .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-v0udIuTaje09TKuj .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-v0udIuTaje09TKuj .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-v0udIuTaje09TKuj :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

简单问题

复杂问题

多步推理

不满意

满意

用户问题

智能路由

直接检索

问题拆解

迭代检索

重排序

多路检索

质量评估

返回答案

专业术语解释：

• 智能路由：根据问题类型自动选择最佳检索策略
• 问题拆解：将复杂问题分解为多个子问题分别检索
• 迭代检索：利用上轮检索结果改进下轮检索
• 自我评估：系统自动评估检索质量并优化

1.6 RAG 全链路增强路线

二、检索前增强

检索前增强的目标是改进用户的查询，使其更容易被检索系统理解。

2.1 HyDE (Hypothetical Document Embeddings)

#mermaid-svg-tKdW5iiWpzpVzGkG{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-tKdW5iiWpzpVzGkG .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-tKdW5iiWpzpVzGkG .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-tKdW5iiWpzpVzGkG .error-icon{fill:#552222;}#mermaid-svg-tKdW5iiWpzpVzGkG .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-tKdW5iiWpzpVzGkG .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-tKdW5iiWpzpVzGkG .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-tKdW5iiWpzpVzGkG .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-tKdW5iiWpzpVzGkG .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-tKdW5iiWpzpVzGkG .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-tKdW5iiWpzpVzGkG .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-tKdW5iiWpzpVzGkG .marker{fill:#333333;stroke:#333333;}#mermaid-svg-tKdW5iiWpzpVzGkG .marker.cross{stroke:#333333;}#mermaid-svg-tKdW5iiWpzpVzGkG svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-tKdW5iiWpzpVzGkG p{margin:0;}#mermaid-svg-tKdW5iiWpzpVzGkG .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-tKdW5iiWpzpVzGkG .cluster-label text{fill:#333;}#mermaid-svg-tKdW5iiWpzpVzGkG .cluster-label span{color:#333;}#mermaid-svg-tKdW5iiWpzpVzGkG .cluster-label span p{background-color:transparent;}#mermaid-svg-tKdW5iiWpzpVzGkG .label text,#mermaid-svg-tKdW5iiWpzpVzGkG span{fill:#333;color:#333;}#mermaid-svg-tKdW5iiWpzpVzGkG .node rect,#mermaid-svg-tKdW5iiWpzpVzGkG .node circle,#mermaid-svg-tKdW5iiWpzpVzGkG .node ellipse,#mermaid-svg-tKdW5iiWpzpVzGkG .node polygon,#mermaid-svg-tKdW5iiWpzpVzGkG .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-tKdW5iiWpzpVzGkG .rough-node .label text,#mermaid-svg-tKdW5iiWpzpVzGkG .node .label text,#mermaid-svg-tKdW5iiWpzpVzGkG .image-shape .label,#mermaid-svg-tKdW5iiWpzpVzGkG .icon-shape .label{text-anchor:middle;}#mermaid-svg-tKdW5iiWpzpVzGkG .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-tKdW5iiWpzpVzGkG .rough-node .label,#mermaid-svg-tKdW5iiWpzpVzGkG .node .label,#mermaid-svg-tKdW5iiWpzpVzGkG .image-shape .label,#mermaid-svg-tKdW5iiWpzpVzGkG .icon-shape .label{text-align:center;}#mermaid-svg-tKdW5iiWpzpVzGkG .node.clickable{cursor:pointer;}#mermaid-svg-tKdW5iiWpzpVzGkG .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-tKdW5iiWpzpVzGkG .arrowheadPath{fill:#333333;}#mermaid-svg-tKdW5iiWpzpVzGkG .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-tKdW5iiWpzpVzGkG .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-tKdW5iiWpzpVzGkG .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-tKdW5iiWpzpVzGkG .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-tKdW5iiWpzpVzGkG .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-tKdW5iiWpzpVzGkG .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-tKdW5iiWpzpVzGkG .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-tKdW5iiWpzpVzGkG .cluster text{fill:#333;}#mermaid-svg-tKdW5iiWpzpVzGkG .cluster span{color:#333;}#mermaid-svg-tKdW5iiWpzpVzGkG div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-tKdW5iiWpzpVzGkG .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-tKdW5iiWpzpVzGkG rect.text{fill:none;stroke-width:0;}#mermaid-svg-tKdW5iiWpzpVzGkG .icon-shape,#mermaid-svg-tKdW5iiWpzpVzGkG .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-tKdW5iiWpzpVzGkG .icon-shape p,#mermaid-svg-tKdW5iiWpzpVzGkG .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-tKdW5iiWpzpVzGkG .icon-shape rect,#mermaid-svg-tKdW5iiWpzpVzGkG .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-tKdW5iiWpzpVzGkG .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-tKdW5iiWpzpVzGkG .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-tKdW5iiWpzpVzGkG :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

用户问题

LLM生成假设答案

假设性文档

向量编码

向量检索

相关文档

返回结果

核心原理：用户的问题往往不够精准。HyDE通过生成假设性答案来改进检索。

工作原理

• 原问题：“手机电池不耐用的原因”
• 生成假设答案：“手机电池不耐用的主要原因包括：后台应用耗电、屏幕亮度设置过高、电池老化…”
• 用假设答案检索：精度大幅提升！

专业术语解释：

• 假设性文档：让LLM生成一个可能的答案，这个答案包含问题相关的关键词和语义
• 向量表示：假设答案的向量表示与真实答案的向量表示更接近

from langchain.prompts import PromptTemplate

# HyDE Prompt
hyde_prompt = PromptTemplate(
input_variables=["question"],
template="""请写一个可能的答案来回答这个问题：{question}

要求：答案应该像是在百科全书或专业文档中找到的那样。"""
)

def hyde_retrieval(question, vectorstore, top_k=3):
# 生成假设性文档
hypothetical_doc = llm(hyde_prompt.format(question=question))

# 用假设文档进行向量检索
retriever = vectorstore.as_retriever(search_kwargs={"k": top_k})
return retriever.get_relevant_documents(hypothetical_doc)

2.2 Query2Doc

#mermaid-svg-kxjd84eVN5T4JhZ4{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-kxjd84eVN5T4JhZ4 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-kxjd84eVN5T4JhZ4 .error-icon{fill:#552222;}#mermaid-svg-kxjd84eVN5T4JhZ4 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-kxjd84eVN5T4JhZ4 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-kxjd84eVN5T4JhZ4 .marker.cross{stroke:#333333;}#mermaid-svg-kxjd84eVN5T4JhZ4 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-kxjd84eVN5T4JhZ4 p{margin:0;}#mermaid-svg-kxjd84eVN5T4JhZ4 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-kxjd84eVN5T4JhZ4 .cluster-label text{fill:#333;}#mermaid-svg-kxjd84eVN5T4JhZ4 .cluster-label span{color:#333;}#mermaid-svg-kxjd84eVN5T4JhZ4 .cluster-label span p{background-color:transparent;}#mermaid-svg-kxjd84eVN5T4JhZ4 .label text,#mermaid-svg-kxjd84eVN5T4JhZ4 span{fill:#333;color:#333;}#mermaid-svg-kxjd84eVN5T4JhZ4 .node rect,#mermaid-svg-kxjd84eVN5T4JhZ4 .node circle,#mermaid-svg-kxjd84eVN5T4JhZ4 .node ellipse,#mermaid-svg-kxjd84eVN5T4JhZ4 .node polygon,#mermaid-svg-kxjd84eVN5T4JhZ4 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-kxjd84eVN5T4JhZ4 .rough-node .label text,#mermaid-svg-kxjd84eVN5T4JhZ4 .node .label text,#mermaid-svg-kxjd84eVN5T4JhZ4 .image-shape .label,#mermaid-svg-kxjd84eVN5T4JhZ4 .icon-shape .label{text-anchor:middle;}#mermaid-svg-kxjd84eVN5T4JhZ4 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-kxjd84eVN5T4JhZ4 .rough-node .label,#mermaid-svg-kxjd84eVN5T4JhZ4 .node .label,#mermaid-svg-kxjd84eVN5T4JhZ4 .image-shape .label,#mermaid-svg-kxjd84eVN5T4JhZ4 .icon-shape .label{text-align:center;}#mermaid-svg-kxjd84eVN5T4JhZ4 .node.clickable{cursor:pointer;}#mermaid-svg-kxjd84eVN5T4JhZ4 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-kxjd84eVN5T4JhZ4 .arrowheadPath{fill:#333333;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-kxjd84eVN5T4JhZ4 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-kxjd84eVN5T4JhZ4 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-kxjd84eVN5T4JhZ4 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-kxjd84eVN5T4JhZ4 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-kxjd84eVN5T4JhZ4 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-kxjd84eVN5T4JhZ4 .cluster text{fill:#333;}#mermaid-svg-kxjd84eVN5T4JhZ4 .cluster span{color:#333;}#mermaid-svg-kxjd84eVN5T4JhZ4 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-kxjd84eVN5T4JhZ4 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-kxjd84eVN5T4JhZ4 rect.text{fill:none;stroke-width:0;}#mermaid-svg-kxjd84eVN5T4JhZ4 .icon-shape,#mermaid-svg-kxjd84eVN5T4JhZ4 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-kxjd84eVN5T4JhZ4 .icon-shape p,#mermaid-svg-kxjd84eVN5T4JhZ4 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-kxjd84eVN5T4JhZ4 .icon-shape rect,#mermaid-svg-kxjd84eVN5T4JhZ4 .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-kxjd84eVN5T4JhZ4 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-kxjd84eVN5T4JhZ4 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-kxjd84eVN5T4JhZ4 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

用户问题

LLM生成文档片段

文档片段50-100字

向量编码

向量检索

相关文档

返回结果

核心原理：与HyDE类似，但Query2Doc专注于生成更短的文档片段来改进检索。

Query2Doc vs HyDE

工作原理：

# Query2Doc Prompt
query2doc_prompt = PromptTemplate(
input_variables=["question"],
template="""基于以下问题，生成一个简短的文档片段（50-100字）。

要求：
1. 包含问题中的关键概念
2. 使用专业术语
3. 保持简洁

问题: {question}

文档片段:"""
)

def query2doc_retrieval(question, vectorstore):
# 生成文档片段
doc_fragment = llm(query2doc_prompt.format(question=question))

# 用文档片段进行检索
retriever = vectorstore.as_retriever(search_kwargs={"k": 3})
return retriever.get_relevant_documents(doc_fragment)

2.3 查询重写 (Query Rewriting)

核心原理：将口语化、模糊的用户问题改写为专业、精确的形式。

# 查询重写Prompt
rewrite_prompt = PromptTemplate(
template="""将以下用户问题改写为更适合搜索的形式。
要求：保留核心意图，使用更专业的术语，补充可能的同义词。

原问题: {question}

改写后的问题:""",
input_variables=["question"]
)

# 示例
# 原问题: "手机老没电咋办"
# 改写后: "手机电池续航时间短，如何解决电池消耗过快的问题"

2.4 查询分解 (Query Decomposition)

核心原理：将复杂的复合问题分解为多个简单的子问题，分别检索后再合并结果。

为什么需要查询分解？

用户问题：“RAG技术有哪些优势？在实际项目中如何应用？有哪些挑战？”

这个问题包含3个子问题，如果直接检索可能找不到同时涵盖所有内容的文档。

解决方案：

专业术语解释：

• 复合问题：包含多个子问题或多个方面的问题
• 子问题独立检索：每个子问题单独进行检索，提升召回率

from langchain.prompts import PromptTemplate

# 查询分解Prompt
decompose_prompt = PromptTemplate(
template="""将以下复杂问题分解为多个独立的子问题。

要求:
1. 每个子问题都应该能独立检索
2. 子问题之间尽量不重复
3. 分解3-5个子问题
4. 每个子问题独立成行

原始问题: {question}

子问题列表（每行一个）:""",
input_variables=["question"]
)

def decompose_query(question, vectorstore):
"""查询分解并检索"""
# 1. 分解问题
sub_questions_str = llm(decompose_prompt.format(question=question))
sub_questions = [q.strip() for q in sub_questions_str.split('\\n') if q.strip()]

print(f"分解出 {len(sub_questions)} 个子问题:")
for i, sq in enumerate(sub_questions, 1):
print(f" {i}. {sq}")

# 2. 对每个子问题检索
all_docs = []
retriever = vectorstore.as_retriever(search_kwargs={"k": 2})

for sq in sub_questions:
docs = retriever.get_relevant_documents(sq)
all_docs.extend(docs)

# 3. 去重并返回Top-N
unique_docs = list({doc.page_content: doc for doc in all_docs}.values())
return unique_docs[:5]

# 使用示例
complex_question = "RAG技术有哪些优势？在实际项目中如何应用？"
docs = decompose_query(complex_question, vectorstore)

2.5 Take Step Back

核心原理：让AI"退一步"思考，从更高层次的概念来理解问题，然后再进行检索。

Take Step Back 示例

用户问题：“Python中的装饰器是怎么工作的？在异步编程中如何使用？”

Step Back后：“Python编程中的装饰器模式和异步编程概念”

优势：从更高层次理解问题的核心概念，检索更全面

专业术语解释：

• 抽象层次：从具体问题上升到更高层次的概念
• 概念关联：通过高层次概念关联更多相关文档

# Take Step Back Prompt
step_back_prompt = PromptTemplate(
template="""你是一个擅长概念抽象的AI助手。

请将以下具体问题抽象为一个更高层次的概念性问题。

要求：
1. 提取问题中的核心概念
2. 从更高层次重新表述问题
3. 保持问题的语义范围

原始问题: {question}

抽象后的问题:""",
input_variables=["question"]
)

def take_step_back_retrieval(question, vectorstore):
"""Step Back检索"""
# 1. 生成抽象问题
abstract_question = llm(step_back_prompt.format(question=question))
print(f"原始问题: {question}")
print(f"抽象问题: {abstract_question}")

# 2. 用抽象问题检索
retriever = vectorstore.as_retriever(search_kwargs={"k": 3})
abstract_docs = retriever.get_relevant_documents(abstract_question)

# 3. 用原始问题检索
original_docs = retriever.get_relevant_documents(question)

# 4. 合并去重
all_docs = abstract_docs + original_docs
unique_docs = list({doc.page_content: doc for doc in all_docs}.values())

return unique_docs[:5]

# 使用示例
question = "Python中的装饰器是怎么工作的？在异步编程中如何使用？"
docs = take_step_back_retrieval(question, vectorstore)

Take Step Back vs 查询分解：

三、检索增强

检索增强的目标是使用多种索引和检索策略，提升检索的召回率和准确率。

3.1 核心概念

多索引检索策略的核心思想是：不同的索引维度捕获不同的信息特征。

#mermaid-svg-xVSYOQeEx7E2tROM{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-xVSYOQeEx7E2tROM .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-xVSYOQeEx7E2tROM .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-xVSYOQeEx7E2tROM .error-icon{fill:#552222;}#mermaid-svg-xVSYOQeEx7E2tROM .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-xVSYOQeEx7E2tROM .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-xVSYOQeEx7E2tROM .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-xVSYOQeEx7E2tROM .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-xVSYOQeEx7E2tROM .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-xVSYOQeEx7E2tROM .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-xVSYOQeEx7E2tROM .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-xVSYOQeEx7E2tROM .marker{fill:#333333;stroke:#333333;}#mermaid-svg-xVSYOQeEx7E2tROM .marker.cross{stroke:#333333;}#mermaid-svg-xVSYOQeEx7E2tROM svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-xVSYOQeEx7E2tROM p{margin:0;}#mermaid-svg-xVSYOQeEx7E2tROM .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-xVSYOQeEx7E2tROM .cluster-label text{fill:#333;}#mermaid-svg-xVSYOQeEx7E2tROM .cluster-label span{color:#333;}#mermaid-svg-xVSYOQeEx7E2tROM .cluster-label span p{background-color:transparent;}#mermaid-svg-xVSYOQeEx7E2tROM .label text,#mermaid-svg-xVSYOQeEx7E2tROM span{fill:#333;color:#333;}#mermaid-svg-xVSYOQeEx7E2tROM .node rect,#mermaid-svg-xVSYOQeEx7E2tROM .node circle,#mermaid-svg-xVSYOQeEx7E2tROM .node ellipse,#mermaid-svg-xVSYOQeEx7E2tROM .node polygon,#mermaid-svg-xVSYOQeEx7E2tROM .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-xVSYOQeEx7E2tROM .rough-node .label text,#mermaid-svg-xVSYOQeEx7E2tROM .node .label text,#mermaid-svg-xVSYOQeEx7E2tROM .image-shape .label,#mermaid-svg-xVSYOQeEx7E2tROM .icon-shape .label{text-anchor:middle;}#mermaid-svg-xVSYOQeEx7E2tROM .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-xVSYOQeEx7E2tROM .rough-node .label,#mermaid-svg-xVSYOQeEx7E2tROM .node .label,#mermaid-svg-xVSYOQeEx7E2tROM .image-shape .label,#mermaid-svg-xVSYOQeEx7E2tROM .icon-shape .label{text-align:center;}#mermaid-svg-xVSYOQeEx7E2tROM .node.clickable{cursor:pointer;}#mermaid-svg-xVSYOQeEx7E2tROM .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-xVSYOQeEx7E2tROM .arrowheadPath{fill:#333333;}#mermaid-svg-xVSYOQeEx7E2tROM .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-xVSYOQeEx7E2tROM .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-xVSYOQeEx7E2tROM .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-xVSYOQeEx7E2tROM .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-xVSYOQeEx7E2tROM .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-xVSYOQeEx7E2tROM .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-xVSYOQeEx7E2tROM .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-xVSYOQeEx7E2tROM .cluster text{fill:#333;}#mermaid-svg-xVSYOQeEx7E2tROM .cluster span{color:#333;}#mermaid-svg-xVSYOQeEx7E2tROM div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-xVSYOQeEx7E2tROM .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-xVSYOQeEx7E2tROM rect.text{fill:none;stroke-width:0;}#mermaid-svg-xVSYOQeEx7E2tROM .icon-shape,#mermaid-svg-xVSYOQeEx7E2tROM .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-xVSYOQeEx7E2tROM .icon-shape p,#mermaid-svg-xVSYOQeEx7E2tROM .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-xVSYOQeEx7E2tROM .icon-shape rect,#mermaid-svg-xVSYOQeEx7E2tROM .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-xVSYOQeEx7E2tROM .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-xVSYOQeEx7E2tROM .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-xVSYOQeEx7E2tROM :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

原始文档

句子索引精准匹配

摘要索引全局视角

问题索引问答匹配

父文档索引上下文完整

多路召回

重排序

最终结果

3.2 混合检索 (Hybrid Search)

核心原理：结合稀疏检索和密集检索的优势。

专业术语解释：

BM25算法原理

BM25是最成功的稀疏检索算法，其核心思想是对词频进行饱和处理：

score

(

D

,

Q

)

=

∑

i

n

IDF

(

q

i

)

×

f

(

q

i

,

D

)

×

(

k

1

+

1

)

f

(

q

i

,

D

)

+

k

1

×

(

1

−

b

+

b

×

∣

D

∣

avgdl

)

\\text{score}(D, Q) = \\sum_{i}^{n} \\text{IDF}(q_i) \\times \\frac{f(q_i, D) \\times (k_1 + 1)}{f(q_i, D) + k_1 \\times (1 – b + b \\times \\frac{|D|}{\\text{avgdl}})}

score(D,Q)=i∑n​IDF(qi​)×f(qi​,D)+k1​×(1−b+b×avgdl∣D∣​)f(qi​,D)×(k1​+1)​

公式解读：

• $f(q_i,D)$
• $|D|$
• $\text{avgdl}$
• $k_1$
• $b$

BM25改进：对TF-IDF的改进在于词频饱和和文档长度归一化，避免长文档占优势。

向量检索原理

Embedding模型：将文本转换为高维向量（通常768-3072维），相似的文本在向量空间中距离更近。

相似度计算：通常使用余弦相似度

similarity

=

cos

⁡

(

θ

)

=

A

⋅

B

∣

∣

A

∣

∣

×

∣

∣

B

∣

∣

\\text{similarity} = \\cos(\\theta) = \\frac{A \\cdot B}{||A|| \\times ||B||}

similarity=cos(θ)=∣∣A∣∣×∣∣B∣∣A⋅B​

RRF融合算法

RRF (Reciprocal Rank Fusion)：将多个检索系统的结果融合，核心公式：

final_score

(

d

)

=

∑

i

=

1

n

1

k

+

rank

i

(

d

)

\\text{final\\_score}(d) = \\sum_{i=1}^{n} \\frac{1}{k + \\text{rank}_i(d)}

final_score(d)=i=1∑n​k+ranki​(d)1​

其中：

• $k$
• ${\text{rank}}_i(d)$

计算示例：

• 排名第1的文档：$\frac{1}{60+0}=0.0167$
• 排名第2的文档：$\frac{1}{60+1}=0.0164$
• 排名第10的文档：$\frac{1}{60+9}=0.0149$

from langchain.retrievers import EnsembleRetriever
from langchain_community.retrievers import BM25Retriever

# 稀疏检索器 (BM25)
bm25_retriever = BM25Retriever.from_documents(documents, k=5)

# 密集检索器 (向量)
vector_retriever = vectorstore.as_retriever(search_kwargs={"k": 5})

# 融合检索器 – 使用RRF算法
ensemble_retriever = EnsembleRetriever(
retrievers=[bm25_retriever, vector_retriever],
weights=[0.5, 0.5] # 可调整权重
)

results = ensemble_retriever.get_relevant_documents("什么是RAG？")

3.3 句子窗口检索 (Sentence Window Retriever)

核心概念：句子窗口检索是一种平衡精度和上下文的技术。

工作原理：

优势：

• 检索时用小窗口，精度高（向量更精准）
• 返回时用大窗口，上下文完整

from llama_index.node_parser import SentenceWindowNodeParser
from llama_index import VectorStoreIndex, Document
from llama_index.postprocessor import MetadataReplacementPostprocessor

# 创建句子窗口解析器
node_parser = SentenceWindowNodeParser.from_defaults(
window_size=3, # 窗口大小：前后各3句
window_metadata_key="window",
original_text_metadata_key="original_text"
)

# 解析文档为节点
nodes = node_parser.get_nodes_from_documents(documents)

# 创建索引
index = VectorStoreIndex(nodes)

# 创建查询引擎（带窗口恢复）
query_engine = index.as_query_engine(
similarity_top_k=2,
node_postprocessors=[
MetadataReplacementPostprocessor(target_metadata_key="window")
]
)

response = query_engine.query("RAG有哪些优化方式？")

3.4 父子文档检索 (Parent Document Retriever)

核心概念：用小块检索，返回大块。

设计思想

#mermaid-svg-eqBdhxYmxw3HeyM6{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-eqBdhxYmxw3HeyM6 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-eqBdhxYmxw3HeyM6 .error-icon{fill:#552222;}#mermaid-svg-eqBdhxYmxw3HeyM6 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-eqBdhxYmxw3HeyM6 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-eqBdhxYmxw3HeyM6 .marker.cross{stroke:#333333;}#mermaid-svg-eqBdhxYmxw3HeyM6 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-eqBdhxYmxw3HeyM6 p{margin:0;}#mermaid-svg-eqBdhxYmxw3HeyM6 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-eqBdhxYmxw3HeyM6 .cluster-label text{fill:#333;}#mermaid-svg-eqBdhxYmxw3HeyM6 .cluster-label span{color:#333;}#mermaid-svg-eqBdhxYmxw3HeyM6 .cluster-label span p{background-color:transparent;}#mermaid-svg-eqBdhxYmxw3HeyM6 .label text,#mermaid-svg-eqBdhxYmxw3HeyM6 span{fill:#333;color:#333;}#mermaid-svg-eqBdhxYmxw3HeyM6 .node rect,#mermaid-svg-eqBdhxYmxw3HeyM6 .node circle,#mermaid-svg-eqBdhxYmxw3HeyM6 .node ellipse,#mermaid-svg-eqBdhxYmxw3HeyM6 .node polygon,#mermaid-svg-eqBdhxYmxw3HeyM6 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-eqBdhxYmxw3HeyM6 .rough-node .label text,#mermaid-svg-eqBdhxYmxw3HeyM6 .node .label text,#mermaid-svg-eqBdhxYmxw3HeyM6 .image-shape .label,#mermaid-svg-eqBdhxYmxw3HeyM6 .icon-shape .label{text-anchor:middle;}#mermaid-svg-eqBdhxYmxw3HeyM6 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-eqBdhxYmxw3HeyM6 .rough-node .label,#mermaid-svg-eqBdhxYmxw3HeyM6 .node .label,#mermaid-svg-eqBdhxYmxw3HeyM6 .image-shape .label,#mermaid-svg-eqBdhxYmxw3HeyM6 .icon-shape .label{text-align:center;}#mermaid-svg-eqBdhxYmxw3HeyM6 .node.clickable{cursor:pointer;}#mermaid-svg-eqBdhxYmxw3HeyM6 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-eqBdhxYmxw3HeyM6 .arrowheadPath{fill:#333333;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-eqBdhxYmxw3HeyM6 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-eqBdhxYmxw3HeyM6 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-eqBdhxYmxw3HeyM6 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-eqBdhxYmxw3HeyM6 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-eqBdhxYmxw3HeyM6 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-eqBdhxYmxw3HeyM6 .cluster text{fill:#333;}#mermaid-svg-eqBdhxYmxw3HeyM6 .cluster span{color:#333;}#mermaid-svg-eqBdhxYmxw3HeyM6 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-eqBdhxYmxw3HeyM6 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-eqBdhxYmxw3HeyM6 rect.text{fill:none;stroke-width:0;}#mermaid-svg-eqBdhxYmxw3HeyM6 .icon-shape,#mermaid-svg-eqBdhxYmxw3HeyM6 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-eqBdhxYmxw3HeyM6 .icon-shape p,#mermaid-svg-eqBdhxYmxw3HeyM6 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-eqBdhxYmxw3HeyM6 .icon-shape rect,#mermaid-svg-eqBdhxYmxw3HeyM6 .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-eqBdhxYmxw3HeyM6 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-eqBdhxYmxw3HeyM6 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-eqBdhxYmxw3HeyM6 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

大文档 1000字

分割

小文档1200字

小文档2200字

小文档3200字

向量检索

匹配到小文档2

返回父文档完整1000字

工作原理：

专业术语解释：

• 父文档：较大的文档块（如1000字），包含完整的上下文
• 子文档：较小的文档块（如200字），用于精准检索

from langchain.retrievers import ParentDocumentRetriever
from langchain.storage import InMemoryStore
from langchain.text_splitter import RecursiveCharacterTextSplitter

# 小文档分割器（用于检索）
child_splitter = RecursiveCharacterTextSplitter(
chunk_size=200, # 200字符一个块
chunk_overlap=50, # 重叠50字符
separators=["\\n\\n", "\\n", "。", "！", "？", "，", " ", ""]
)

# 大文档分割器（用于返回）
parent_splitter = RecursiveCharacterTextSplitter(
chunk_size=1000, # 1000字符一个块
chunk_overlap=100,
separators=["\\n\\n", "\\n", "。", "！", "？"]
)

# 创建父文档检索器
retriever = ParentDocumentRetriever(
vectorstore=vectorstore,
docstore=store,
child_splitter=child_splitter,
parent_splitter=parent_splitter
)

3.5 摘要检索 (Summary Retriever)

#mermaid-svg-dsknaoLZTw9N2qX6{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-dsknaoLZTw9N2qX6 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-dsknaoLZTw9N2qX6 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-dsknaoLZTw9N2qX6 .error-icon{fill:#552222;}#mermaid-svg-dsknaoLZTw9N2qX6 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-dsknaoLZTw9N2qX6 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-dsknaoLZTw9N2qX6 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-dsknaoLZTw9N2qX6 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-dsknaoLZTw9N2qX6 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-dsknaoLZTw9N2qX6 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-dsknaoLZTw9N2qX6 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-dsknaoLZTw9N2qX6 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-dsknaoLZTw9N2qX6 .marker.cross{stroke:#333333;}#mermaid-svg-dsknaoLZTw9N2qX6 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-dsknaoLZTw9N2qX6 p{margin:0;}#mermaid-svg-dsknaoLZTw9N2qX6 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-dsknaoLZTw9N2qX6 .cluster-label text{fill:#333;}#mermaid-svg-dsknaoLZTw9N2qX6 .cluster-label span{color:#333;}#mermaid-svg-dsknaoLZTw9N2qX6 .cluster-label span p{background-color:transparent;}#mermaid-svg-dsknaoLZTw9N2qX6 .label text,#mermaid-svg-dsknaoLZTw9N2qX6 span{fill:#333;color:#333;}#mermaid-svg-dsknaoLZTw9N2qX6 .node rect,#mermaid-svg-dsknaoLZTw9N2qX6 .node circle,#mermaid-svg-dsknaoLZTw9N2qX6 .node ellipse,#mermaid-svg-dsknaoLZTw9N2qX6 .node polygon,#mermaid-svg-dsknaoLZTw9N2qX6 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-dsknaoLZTw9N2qX6 .rough-node .label text,#mermaid-svg-dsknaoLZTw9N2qX6 .node .label text,#mermaid-svg-dsknaoLZTw9N2qX6 .image-shape .label,#mermaid-svg-dsknaoLZTw9N2qX6 .icon-shape .label{text-anchor:middle;}#mermaid-svg-dsknaoLZTw9N2qX6 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-dsknaoLZTw9N2qX6 .rough-node .label,#mermaid-svg-dsknaoLZTw9N2qX6 .node .label,#mermaid-svg-dsknaoLZTw9N2qX6 .image-shape .label,#mermaid-svg-dsknaoLZTw9N2qX6 .icon-shape .label{text-align:center;}#mermaid-svg-dsknaoLZTw9N2qX6 .node.clickable{cursor:pointer;}#mermaid-svg-dsknaoLZTw9N2qX6 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-dsknaoLZTw9N2qX6 .arrowheadPath{fill:#333333;}#mermaid-svg-dsknaoLZTw9N2qX6 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-dsknaoLZTw9N2qX6 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-dsknaoLZTw9N2qX6 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-dsknaoLZTw9N2qX6 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-dsknaoLZTw9N2qX6 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-dsknaoLZTw9N2qX6 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-dsknaoLZTw9N2qX6 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-dsknaoLZTw9N2qX6 .cluster text{fill:#333;}#mermaid-svg-dsknaoLZTw9N2qX6 .cluster span{color:#333;}#mermaid-svg-dsknaoLZTw9N2qX6 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-dsknaoLZTw9N2qX6 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-dsknaoLZTw9N2qX6 rect.text{fill:none;stroke-width:0;}#mermaid-svg-dsknaoLZTw9N2qX6 .icon-shape,#mermaid-svg-dsknaoLZTw9N2qX6 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-dsknaoLZTw9N2qX6 .icon-shape p,#mermaid-svg-dsknaoLZTw9N2qX6 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-dsknaoLZTw9N2qX6 .icon-shape rect,#mermaid-svg-dsknaoLZTw9N2qX6 .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-dsknaoLZTw9N2qX6 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-dsknaoLZTw9N2qX6 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-dsknaoLZTw9N2qX6 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

原始文档5000字

生成摘要200字

摘要索引向量存储

用户问题

向量检索

匹配摘要

返回完整文档

核心概念：用文档摘要进行检索，返回完整文档。

为什么需要摘要检索？

场景：用户问"RAG技术在企业中的应用"

问题：

• 原始文档太长（5000字），向量检索容易聚焦细节而忽略主题
• 摘要概括了文档的核心内容（200字）

解决方案：

• 用摘要向量进行检索（匹配主题）
• 返回完整文档（提供细节）

工作原理：

专业术语解释：

• 摘要索引：存储文档摘要的向量索引，用于主题级别的检索
• 全局视角：摘要提供文档的全局视图，避免陷入细节

from langchain.prompts import PromptTemplate
from langchain.llms import OpenAI
from langchain.schema import Document

# 1. 摘要生成Prompt
summary_prompt = PromptTemplate(
template="""请为以下文档生成一个简洁的摘要（100字以内）：

文档:
{context}

摘要:""",
input_variables=["context"]
)

# 2. 为文档生成摘要
def generate_summary(documents):
llm = OpenAI(temperature=0)
summaries = []

for i, doc in enumerate(documents):
print(f"处理第 {i+1}/{len(documents)} 个文档…")

# 生成摘要
summary = llm(summary_prompt.format(context=doc.page_content))

summaries.append({
"summary": summary,
"original_doc": doc
})

return summaries

# 3. 为知识库生成摘要
summaries = generate_summary(documents)

# 4. 用摘要构建索引
summary_docs = [
Document(
page_content=s["summary"],
metadata={"doc_id": i}
)
for i, s in enumerate(summaries)
]

summary_index = Chroma.from_documents(
summary_docs,
embeddings,
collection_name="summary_index"
)

# 5. 检索函数
def retrieve_by_summary(query):
"""通过摘要检索，返回原始文档"""
# 找到相似的摘要
similar_summaries = summary_index.similarity_search(query, k=1)

if not similar_summaries:
return []

# 获取对应的原始文档
doc_id = similar_summaries[0].metadata["doc_id"]
return [summaries[doc_id]["original_doc"]]

# 使用示例
results = retrieve_by_summary("RAG在企业中的应用场景")

3.6 假设性问题检索 (Hypothetical Questions Retriever)

#mermaid-svg-VhHYhUZxffWjKkhh{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-VhHYhUZxffWjKkhh .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-VhHYhUZxffWjKkhh .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-VhHYhUZxffWjKkhh .error-icon{fill:#552222;}#mermaid-svg-VhHYhUZxffWjKkhh .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-VhHYhUZxffWjKkhh .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-VhHYhUZxffWjKkhh .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-VhHYhUZxffWjKkhh .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-VhHYhUZxffWjKkhh .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-VhHYhUZxffWjKkhh .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-VhHYhUZxffWjKkhh .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-VhHYhUZxffWjKkhh .marker{fill:#333333;stroke:#333333;}#mermaid-svg-VhHYhUZxffWjKkhh .marker.cross{stroke:#333333;}#mermaid-svg-VhHYhUZxffWjKkhh svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-VhHYhUZxffWjKkhh p{margin:0;}#mermaid-svg-VhHYhUZxffWjKkhh .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-VhHYhUZxffWjKkhh .cluster-label text{fill:#333;}#mermaid-svg-VhHYhUZxffWjKkhh .cluster-label span{color:#333;}#mermaid-svg-VhHYhUZxffWjKkhh .cluster-label span p{background-color:transparent;}#mermaid-svg-VhHYhUZxffWjKkhh .label text,#mermaid-svg-VhHYhUZxffWjKkhh span{fill:#333;color:#333;}#mermaid-svg-VhHYhUZxffWjKkhh .node rect,#mermaid-svg-VhHYhUZxffWjKkhh .node circle,#mermaid-svg-VhHYhUZxffWjKkhh .node ellipse,#mermaid-svg-VhHYhUZxffWjKkhh .node polygon,#mermaid-svg-VhHYhUZxffWjKkhh .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-VhHYhUZxffWjKkhh .rough-node .label text,#mermaid-svg-VhHYhUZxffWjKkhh .node .label text,#mermaid-svg-VhHYhUZxffWjKkhh .image-shape .label,#mermaid-svg-VhHYhUZxffWjKkhh .icon-shape .label{text-anchor:middle;}#mermaid-svg-VhHYhUZxffWjKkhh .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-VhHYhUZxffWjKkhh .rough-node .label,#mermaid-svg-VhHYhUZxffWjKkhh .node .label,#mermaid-svg-VhHYhUZxffWjKkhh .image-shape .label,#mermaid-svg-VhHYhUZxffWjKkhh .icon-shape .label{text-align:center;}#mermaid-svg-VhHYhUZxffWjKkhh .node.clickable{cursor:pointer;}#mermaid-svg-VhHYhUZxffWjKkhh .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-VhHYhUZxffWjKkhh .arrowheadPath{fill:#333333;}#mermaid-svg-VhHYhUZxffWjKkhh .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-VhHYhUZxffWjKkhh .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-VhHYhUZxffWjKkhh .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-VhHYhUZxffWjKkhh .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-VhHYhUZxffWjKkhh .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-VhHYhUZxffWjKkhh .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-VhHYhUZxffWjKkhh .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-VhHYhUZxffWjKkhh .cluster text{fill:#333;}#mermaid-svg-VhHYhUZxffWjKkhh .cluster span{color:#333;}#mermaid-svg-VhHYhUZxffWjKkhh div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-VhHYhUZxffWjKkhh .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-VhHYhUZxffWjKkhh rect.text{fill:none;stroke-width:0;}#mermaid-svg-VhHYhUZxffWjKkhh .icon-shape,#mermaid-svg-VhHYhUZxffWjKkhh .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-VhHYhUZxffWjKkhh .icon-shape p,#mermaid-svg-VhHYhUZxffWjKkhh .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-VhHYhUZxffWjKkhh .icon-shape rect,#mermaid-svg-VhHYhUZxffWjKkhh .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-VhHYhUZxffWjKkhh .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-VhHYhUZxffWjKkhh .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-VhHYhUZxffWjKkhh :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

原始文档

LLM生成假设问题

问题1Python是什么时候创建的?

问题2Python的创始人是谁?

问题3Python有什么特点?

问题索引向量存储

用户问题Python的历史

向量检索

匹配问题

返回原始文档

核心概念：预先生成文档可能回答的问题，用这些问题进行检索。

应用场景

文档：“Python是一门高级编程语言，由Guido van Rossum于1991年创建…”

生成假设性问题：

• “Python是什么时候创建的？”
• “Python的创始人是谁？”
• “Python有什么特点？”

检索时：用户问"Python的历史" → 匹配到"Python是什么时候创建的？" → 返回原文档

工作原理：

from langchain.prompts import PromptTemplate

# 问题生成Prompt
question_gen_prompt = PromptTemplate(
template="""阅读以下文档，生成3个可能的问题。

要求:
1. 问题应该涵盖文档的核心内容
2. 问题的表达方式应该多样化
3. 每个问题独立成行

文档:
{context}

问题列表:""",
input_variables=["context"]
)

def generate_questions(documents):
"""为文档生成假设性问题"""
llm = OpenAI(temperature=0.7)
all_questions = []

for i, doc in enumerate(documents, 1):
# 生成问题
response = llm(question_gen_prompt.format(context=doc.page_content))
questions = [q.strip() for q in response.split('\\n') if q.strip()]

# 存储问题与文档的映射
for q in questions:
all_questions.append({
"question": q,
"source_doc": doc,
"source_id": i
})

return all_questions

# 使用示例
questions = generate_questions(documents)

# 用问题构建索引
question_docs = [
Document(page_content=q["question"], metadata={"source_id": q["source_id"]})
for q in questions
]

question_index = Chroma.from_documents(question_docs, embeddings)

3.7 多索引融合策略

#mermaid-svg-WzTycUmhToTyUlO9{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-WzTycUmhToTyUlO9 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-WzTycUmhToTyUlO9 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-WzTycUmhToTyUlO9 .error-icon{fill:#552222;}#mermaid-svg-WzTycUmhToTyUlO9 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-WzTycUmhToTyUlO9 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-WzTycUmhToTyUlO9 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-WzTycUmhToTyUlO9 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-WzTycUmhToTyUlO9 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-WzTycUmhToTyUlO9 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-WzTycUmhToTyUlO9 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-WzTycUmhToTyUlO9 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-WzTycUmhToTyUlO9 .marker.cross{stroke:#333333;}#mermaid-svg-WzTycUmhToTyUlO9 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-WzTycUmhToTyUlO9 p{margin:0;}#mermaid-svg-WzTycUmhToTyUlO9 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-WzTycUmhToTyUlO9 .cluster-label text{fill:#333;}#mermaid-svg-WzTycUmhToTyUlO9 .cluster-label span{color:#333;}#mermaid-svg-WzTycUmhToTyUlO9 .cluster-label span p{background-color:transparent;}#mermaid-svg-WzTycUmhToTyUlO9 .label text,#mermaid-svg-WzTycUmhToTyUlO9 span{fill:#333;color:#333;}#mermaid-svg-WzTycUmhToTyUlO9 .node rect,#mermaid-svg-WzTycUmhToTyUlO9 .node circle,#mermaid-svg-WzTycUmhToTyUlO9 .node ellipse,#mermaid-svg-WzTycUmhToTyUlO9 .node polygon,#mermaid-svg-WzTycUmhToTyUlO9 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-WzTycUmhToTyUlO9 .rough-node .label text,#mermaid-svg-WzTycUmhToTyUlO9 .node .label text,#mermaid-svg-WzTycUmhToTyUlO9 .image-shape .label,#mermaid-svg-WzTycUmhToTyUlO9 .icon-shape .label{text-anchor:middle;}#mermaid-svg-WzTycUmhToTyUlO9 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-WzTycUmhToTyUlO9 .rough-node .label,#mermaid-svg-WzTycUmhToTyUlO9 .node .label,#mermaid-svg-WzTycUmhToTyUlO9 .image-shape .label,#mermaid-svg-WzTycUmhToTyUlO9 .icon-shape .label{text-align:center;}#mermaid-svg-WzTycUmhToTyUlO9 .node.clickable{cursor:pointer;}#mermaid-svg-WzTycUmhToTyUlO9 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-WzTycUmhToTyUlO9 .arrowheadPath{fill:#333333;}#mermaid-svg-WzTycUmhToTyUlO9 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-WzTycUmhToTyUlO9 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-WzTycUmhToTyUlO9 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-WzTycUmhToTyUlO9 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-WzTycUmhToTyUlO9 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-WzTycUmhToTyUlO9 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-WzTycUmhToTyUlO9 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-WzTycUmhToTyUlO9 .cluster text{fill:#333;}#mermaid-svg-WzTycUmhToTyUlO9 .cluster span{color:#333;}#mermaid-svg-WzTycUmhToTyUlO9 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-WzTycUmhToTyUlO9 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-WzTycUmhToTyUlO9 rect.text{fill:none;stroke-width:0;}#mermaid-svg-WzTycUmhToTyUlO9 .icon-shape,#mermaid-svg-WzTycUmhToTyUlO9 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-WzTycUmhToTyUlO9 .icon-shape p,#mermaid-svg-WzTycUmhToTyUlO9 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-WzTycUmhToTyUlO9 .icon-shape rect,#mermaid-svg-WzTycUmhToTyUlO9 .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-WzTycUmhToTyUlO9 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-WzTycUmhToTyUlO9 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-WzTycUmhToTyUlO9 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

用户问题

句子索引检索

摘要索引检索

问题索引检索

父文档索引检索

多路召回合并结果

去重移除重复文档

重排序Re-rank

Top-N结果

核心概念：同时使用多种索引策略，融合所有结果。

专业术语解释：

• 多路召回：从多个索引通道同时检索文档
• 结果融合：将多个检索结果合并、去重、排序

def multi_index_retrieval(query):
"""多索引融合检索"""
all_docs = []

# 1. 句子窗口检索
sentence_results = sentence_retriever.get_relevant_documents(query)
all_docs.extend(sentence_results)

# 2. 摘要检索
summary_results = summary_retriever.get_relevant_documents(query)
all_docs.extend(summary_results)

# 3. 问题检索
question_results = question_retriever.get_relevant_documents(query)
all_docs.extend(question_results)

# 4. 去重
unique_docs = list({doc.page_content: doc for doc in all_docs}.values())

# 5. 重排序
reranked_docs = reranker(query, unique_docs)

return reranked_docs[:5]

四、检索后增强

检索后增强的目标是对检索结果进行优化，提升最终答案的质量。

4.1 重排序 (Re-ranking) 原理

核心问题：向量检索使用余弦相似度，但这不等于问题相关性！

例子

用户问：“如何减肥？”

• 文档A（向量相似度0.85）：“减肥需要控制饮食和运动” – 直接相关
• 文档B（向量相似度0.87）：“这个文档讨论了各种健康问题，包括如何增重、如何减肥…” – 相关但啰嗦

重排序后：文档A排第一，因为直接回答问题

专业术语解释：

• 双塔模型：向量检索使用的模型，查询和文档独立编码
• 交叉编码器：重排序使用的模型，查询和文档一起编码，精度更高但速度更慢

#mermaid-svg-6OQAgoxFNffYeUS1{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;fill:#333;}@keyframes edge-animation-frame{from{stroke-dashoffset:0;}}@keyframes dash{to{stroke-dashoffset:0;}}#mermaid-svg-6OQAgoxFNffYeUS1 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-6OQAgoxFNffYeUS1 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-6OQAgoxFNffYeUS1 .error-icon{fill:#552222;}#mermaid-svg-6OQAgoxFNffYeUS1 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-6OQAgoxFNffYeUS1 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-6OQAgoxFNffYeUS1 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-6OQAgoxFNffYeUS1 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-6OQAgoxFNffYeUS1 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-6OQAgoxFNffYeUS1 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-6OQAgoxFNffYeUS1 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-6OQAgoxFNffYeUS1 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-6OQAgoxFNffYeUS1 .marker.cross{stroke:#333333;}#mermaid-svg-6OQAgoxFNffYeUS1 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-6OQAgoxFNffYeUS1 p{margin:0;}#mermaid-svg-6OQAgoxFNffYeUS1 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-6OQAgoxFNffYeUS1 .cluster-label text{fill:#333;}#mermaid-svg-6OQAgoxFNffYeUS1 .cluster-label span{color:#333;}#mermaid-svg-6OQAgoxFNffYeUS1 .cluster-label span p{background-color:transparent;}#mermaid-svg-6OQAgoxFNffYeUS1 .label text,#mermaid-svg-6OQAgoxFNffYeUS1 span{fill:#333;color:#333;}#mermaid-svg-6OQAgoxFNffYeUS1 .node rect,#mermaid-svg-6OQAgoxFNffYeUS1 .node circle,#mermaid-svg-6OQAgoxFNffYeUS1 .node ellipse,#mermaid-svg-6OQAgoxFNffYeUS1 .node polygon,#mermaid-svg-6OQAgoxFNffYeUS1 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-6OQAgoxFNffYeUS1 .rough-node .label text,#mermaid-svg-6OQAgoxFNffYeUS1 .node .label text,#mermaid-svg-6OQAgoxFNffYeUS1 .image-shape .label,#mermaid-svg-6OQAgoxFNffYeUS1 .icon-shape .label{text-anchor:middle;}#mermaid-svg-6OQAgoxFNffYeUS1 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-6OQAgoxFNffYeUS1 .rough-node .label,#mermaid-svg-6OQAgoxFNffYeUS1 .node .label,#mermaid-svg-6OQAgoxFNffYeUS1 .image-shape .label,#mermaid-svg-6OQAgoxFNffYeUS1 .icon-shape .label{text-align:center;}#mermaid-svg-6OQAgoxFNffYeUS1 .node.clickable{cursor:pointer;}#mermaid-svg-6OQAgoxFNffYeUS1 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-6OQAgoxFNffYeUS1 .arrowheadPath{fill:#333333;}#mermaid-svg-6OQAgoxFNffYeUS1 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-6OQAgoxFNffYeUS1 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-6OQAgoxFNffYeUS1 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-6OQAgoxFNffYeUS1 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-6OQAgoxFNffYeUS1 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-6OQAgoxFNffYeUS1 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-6OQAgoxFNffYeUS1 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-6OQAgoxFNffYeUS1 .cluster text{fill:#333;}#mermaid-svg-6OQAgoxFNffYeUS1 .cluster span{color:#333;}#mermaid-svg-6OQAgoxFNffYeUS1 div.mermaidTooltip{position:absolute;text-align:center;max-width:200px;padding:2px;font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:12px;background:hsl(80, 100%, 96.2745098039%);border:1px solid #aaaa33;border-radius:2px;pointer-events:none;z-index:100;}#mermaid-svg-6OQAgoxFNffYeUS1 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-6OQAgoxFNffYeUS1 rect.text{fill:none;stroke-width:0;}#mermaid-svg-6OQAgoxFNffYeUS1 .icon-shape,#mermaid-svg-6OQAgoxFNffYeUS1 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-6OQAgoxFNffYeUS1 .icon-shape p,#mermaid-svg-6OQAgoxFNffYeUS1 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-6OQAgoxFNffYeUS1 .icon-shape rect,#mermaid-svg-6OQAgoxFNffYeUS1 .image-shape rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-6OQAgoxFNffYeUS1 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-6OQAgoxFNffYeUS1 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-6OQAgoxFNffYeUS1 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

向量检索Top-10文档

粗排结果

Re-rank模型精打细算

Top-3高质量文档

4.2 主流Re-ranker对比

4.3 Re-ranking实现

# 方案1: Cohere Rerank
import cohere

def cohere_rerank(query, documents, top_n=3):
co = cohere.Client('your-api-key')
results = co.rerank(
model="rerank-v3.5",
query=query,
documents=[doc.page_content for doc in documents],
top_n=top_n
)
return [documents[r.index] for r in results.results]

# 方案2: BGE Reranker
from sentence_transformers import CrossEncoder

def bge_rerank(query, documents, top_k=3):
model = CrossEncoder('BAAI/bge-reranker-v2-m3')

# 构建查询-文档对
pairs = [[query, doc.page_content] for doc in documents]

# 计算分数
scores = model.predict(pairs)

# 排序
scored_docs = sorted(zip(documents, scores), key=lambda x: x[1], reverse=True)
return [doc for doc, score in scored_docs[:top_k]]

4.4 上下文压缩

核心概念：对检索到的文档进行压缩，去除无关内容，保留关键信息。

专业术语解释：

• 上下文压缩：提取文档中与问题最相关的片段
• 信息密度：单位长度内的有效信息量

from langchain.retrievers import ContextualCompressionRetriever
from langchain.document_compressors import LLMChainExtractor

# 创建压缩器
compressor = LLMChainExtractor.from_llm(ChatOpenAI(model="gpt-4"))

# 创建压缩检索器
compression_retriever = ContextualCompressionRetriever(
base_compressor=compressor,
base_retriever=vectorstore.as_retriever()
)

# 检索并压缩
compressed_docs = compression_retriever.get_relevant_documents("什么是RAG？")

五、完整生产级示例

5.1 LangChain生产级实现

import os
from typing import List, Dict
from langchain.embeddings import OpenAIEmbeddings
from langchain.vectorstores import Chroma
from langchain.chat_models import ChatOpenAI
from langchain.chains import RetrievalQA
from langchain.retrievers import ContextualCompressionRetriever
from langchain_community.document_compressors.cohere_rerank import CohereRerank
from langchain.retrievers import EnsembleRetriever
from langchain_community.retrievers import BM25Retriever

class ProductionRAG:
"""生产级RAG系统"""

def __init__(
self,
knowledge_base: List[str],
model: str = "gpt-4",
top_k: int = 20,
rerank_top_n: int = 5
):
# 初始化嵌入模型
self.embeddings = OpenAIEmbeddings()

# 创建向量数据库
self.vectorstore = Chroma(
collection_name="production_kb",
embedding_function=self.embeddings
)

# 添加文档
if self.vectorstore._collection.count() == 0:
from langchain.schema import Document
docs = [Document(page_content=text) for text in knowledge_base]
self.vectorstore.add_documents(docs)

# 混合检索器
bm25_retriever = BM25Retriever.from_documents(docs, k=top_k)
vector_retriever = self.vectorstore.as_retriever(search_kwargs={"k": top_k})

# 融合检索器
self.retriever = EnsembleRetriever(
retrievers=[bm25_retriever, vector_retriever],
weights=[0.4, 0.6]
)

# 重排序
compressor = CohereRerank(top_n=rerank_top_n)
self.compression_retriever = ContextualCompressionRetriever(
base_compressor=compressor,
base_retriever=self.retriever
)

# 创建RAG链
self.qa_chain = RetrievalQA.from_chain_type(
llm=ChatOpenAI(model=model, temperature=0),
retriever=self.compression_retriever,
return_source_documents=True
)

def query(self, question: str) –> Dict:
"""执行查询"""
result = self.qa_chain({"query": question})
return {
"question": question,
"answer": result["result"],
"sources": result.get("source_documents", [])
}

5.2 LlamaIndex生产级实现

from llama_index import VectorStoreIndex, ServiceContext
from llama_index.llms import OpenAI
from llama_index.embeddings import OpenAIEmbedding
from llama_index.node_parser import SentenceWindowNodeParser
from llama_index.postprocessor import MetadataReplacementPostprocessor

class ProductionRAGLlama:
"""生产级RAG系统 – LlamaIndex版本"""

def __init__(self, data_dir: str = "./data"):
# 配置服务上下文
self.service_context = ServiceContext.from_defaults(
llm=OpenAI(model="gpt-4", temperature=0),
embed_model=OpenAIEmbedding(model="text-embedding-ada-002"),
chunk_size=512,
chunk_overlap=50
)

# 加载文档
from llama_index import SimpleDirectoryReader
documents = SimpleDirectoryReader(data_dir).load_data()

# 句子窗口解析器
node_parser = SentenceWindowNodeParser.from_defaults(window_size=3)
nodes = node_parser.get_nodes_from_documents(documents)

# 创建索引
self.index = VectorStoreIndex(nodes, service_context=self.service_context)

# 创建查询引擎
self.query_engine = self.index.as_query_engine(
similarity_top_k=10,
node_postprocessors=[
MetadataReplacementPostprocessor(target_metadata_key="window")
],
response_mode="compact"
)

def query(self, question: str) –> str:
"""执行查询"""
response = self.query_engine.query(question)
return str(response)