Redis向量检索-171主机测评

Redis 向量检索完全指南：原理、HNSW 算法、实战与选型（2026 最新）

适用技术栈：Redis 8 / Redis Stack、Python（redis-py）、Java（Spring AI / Jedis）适用读者：后端 / AI 应用开发者、RAG 系统设计者、面试备考者

作者：一位想进大厂的树先生：Mr.Tree。

一、什么是向量检索

1.1 从"关键词检索"到"语义检索"

传统检索（如 MySQL LIKE、Elasticsearch 倒排）依赖关键词精确匹配：你搜"苹果手机"，它匹配不到只写了"iPhone"的文档，因为字面不一样。

向量检索（Vector Search）把内容映射成高维空间中的一个点（向量），语义相近的内容在空间中距离也近。"苹果手机"和"iPhone"虽然字面不同，但向量位置很接近，因此能被搜到。

#mermaid-svg-id4G1LHth8io1NNw{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-id4G1LHth8io1NNw .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-id4G1LHth8io1NNw .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-id4G1LHth8io1NNw .error-icon{fill:#552222;}#mermaid-svg-id4G1LHth8io1NNw .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-id4G1LHth8io1NNw .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-id4G1LHth8io1NNw .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-id4G1LHth8io1NNw .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-id4G1LHth8io1NNw .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-id4G1LHth8io1NNw .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-id4G1LHth8io1NNw .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-id4G1LHth8io1NNw .marker{fill:#333333;stroke:#333333;}#mermaid-svg-id4G1LHth8io1NNw .marker.cross{stroke:#333333;}#mermaid-svg-id4G1LHth8io1NNw svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-id4G1LHth8io1NNw p{margin:0;}#mermaid-svg-id4G1LHth8io1NNw .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-id4G1LHth8io1NNw .cluster-label text{fill:#333;}#mermaid-svg-id4G1LHth8io1NNw .cluster-label span{color:#333;}#mermaid-svg-id4G1LHth8io1NNw .cluster-label span p{background-color:transparent;}#mermaid-svg-id4G1LHth8io1NNw .label text,#mermaid-svg-id4G1LHth8io1NNw span{fill:#333;color:#333;}#mermaid-svg-id4G1LHth8io1NNw .node rect,#mermaid-svg-id4G1LHth8io1NNw .node circle,#mermaid-svg-id4G1LHth8io1NNw .node ellipse,#mermaid-svg-id4G1LHth8io1NNw .node polygon,#mermaid-svg-id4G1LHth8io1NNw .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-id4G1LHth8io1NNw .rough-node .label text,#mermaid-svg-id4G1LHth8io1NNw .node .label text,#mermaid-svg-id4G1LHth8io1NNw .image-shape .label,#mermaid-svg-id4G1LHth8io1NNw .icon-shape .label{text-anchor:middle;}#mermaid-svg-id4G1LHth8io1NNw .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-id4G1LHth8io1NNw .rough-node .label,#mermaid-svg-id4G1LHth8io1NNw .node .label,#mermaid-svg-id4G1LHth8io1NNw .image-shape .label,#mermaid-svg-id4G1LHth8io1NNw .icon-shape .label{text-align:center;}#mermaid-svg-id4G1LHth8io1NNw .node.clickable{cursor:pointer;}#mermaid-svg-id4G1LHth8io1NNw .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-id4G1LHth8io1NNw .arrowheadPath{fill:#333333;}#mermaid-svg-id4G1LHth8io1NNw .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-id4G1LHth8io1NNw .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-id4G1LHth8io1NNw .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-id4G1LHth8io1NNw .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-id4G1LHth8io1NNw .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-id4G1LHth8io1NNw .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-id4G1LHth8io1NNw .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-id4G1LHth8io1NNw .cluster text{fill:#333;}#mermaid-svg-id4G1LHth8io1NNw .cluster span{color:#333;}#mermaid-svg-id4G1LHth8io1NNw 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-id4G1LHth8io1NNw .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-id4G1LHth8io1NNw rect.text{fill:none;stroke-width:0;}#mermaid-svg-id4G1LHth8io1NNw .icon-shape,#mermaid-svg-id4G1LHth8io1NNw .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-id4G1LHth8io1NNw .icon-shape p,#mermaid-svg-id4G1LHth8io1NNw .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-id4G1LHth8io1NNw .icon-shape .label rect,#mermaid-svg-id4G1LHth8io1NNw .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-id4G1LHth8io1NNw .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-id4G1LHth8io1NNw .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-id4G1LHth8io1NNw :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

向量检索

查询：苹果手机

语义距离

能匹配：iPhone（语义相近）

关键词检索

查询：苹果手机

字面匹配

匹配不到：iPhone

1.2 什么是向量嵌入（Embedding）

Embedding 是把文本、图片、商品等对象，通过模型转换成一串浮点数（向量）。例如 OpenAI 的 text-embedding 模型输出 1536 维向量，bge 等开源模型常见 768 / 1024 维。

#mermaid-svg-jkDBGRXAmOU6oOB0{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-jkDBGRXAmOU6oOB0 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-jkDBGRXAmOU6oOB0 .error-icon{fill:#552222;}#mermaid-svg-jkDBGRXAmOU6oOB0 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-jkDBGRXAmOU6oOB0 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-jkDBGRXAmOU6oOB0 .marker.cross{stroke:#333333;}#mermaid-svg-jkDBGRXAmOU6oOB0 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-jkDBGRXAmOU6oOB0 p{margin:0;}#mermaid-svg-jkDBGRXAmOU6oOB0 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-jkDBGRXAmOU6oOB0 .cluster-label text{fill:#333;}#mermaid-svg-jkDBGRXAmOU6oOB0 .cluster-label span{color:#333;}#mermaid-svg-jkDBGRXAmOU6oOB0 .cluster-label span p{background-color:transparent;}#mermaid-svg-jkDBGRXAmOU6oOB0 .label text,#mermaid-svg-jkDBGRXAmOU6oOB0 span{fill:#333;color:#333;}#mermaid-svg-jkDBGRXAmOU6oOB0 .node rect,#mermaid-svg-jkDBGRXAmOU6oOB0 .node circle,#mermaid-svg-jkDBGRXAmOU6oOB0 .node ellipse,#mermaid-svg-jkDBGRXAmOU6oOB0 .node polygon,#mermaid-svg-jkDBGRXAmOU6oOB0 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-jkDBGRXAmOU6oOB0 .rough-node .label text,#mermaid-svg-jkDBGRXAmOU6oOB0 .node .label text,#mermaid-svg-jkDBGRXAmOU6oOB0 .image-shape .label,#mermaid-svg-jkDBGRXAmOU6oOB0 .icon-shape .label{text-anchor:middle;}#mermaid-svg-jkDBGRXAmOU6oOB0 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-jkDBGRXAmOU6oOB0 .rough-node .label,#mermaid-svg-jkDBGRXAmOU6oOB0 .node .label,#mermaid-svg-jkDBGRXAmOU6oOB0 .image-shape .label,#mermaid-svg-jkDBGRXAmOU6oOB0 .icon-shape .label{text-align:center;}#mermaid-svg-jkDBGRXAmOU6oOB0 .node.clickable{cursor:pointer;}#mermaid-svg-jkDBGRXAmOU6oOB0 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-jkDBGRXAmOU6oOB0 .arrowheadPath{fill:#333333;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-jkDBGRXAmOU6oOB0 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-jkDBGRXAmOU6oOB0 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-jkDBGRXAmOU6oOB0 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-jkDBGRXAmOU6oOB0 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-jkDBGRXAmOU6oOB0 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-jkDBGRXAmOU6oOB0 .cluster text{fill:#333;}#mermaid-svg-jkDBGRXAmOU6oOB0 .cluster span{color:#333;}#mermaid-svg-jkDBGRXAmOU6oOB0 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-jkDBGRXAmOU6oOB0 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-jkDBGRXAmOU6oOB0 rect.text{fill:none;stroke-width:0;}#mermaid-svg-jkDBGRXAmOU6oOB0 .icon-shape,#mermaid-svg-jkDBGRXAmOU6oOB0 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-jkDBGRXAmOU6oOB0 .icon-shape p,#mermaid-svg-jkDBGRXAmOU6oOB0 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-jkDBGRXAmOU6oOB0 .icon-shape .label rect,#mermaid-svg-jkDBGRXAmOU6oOB0 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-jkDBGRXAmOU6oOB0 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-jkDBGRXAmOU6oOB0 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-jkDBGRXAmOU6oOB0 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

文本/图片/商品

Embedding 模型

高维向量，如 1536 个浮点数

存入 Redis 并建立索引

1.3 典型应用场景

场景说明

RAG（检索增强生成）	先从知识库检索相关文档，再交给大模型生成答案
语义搜索	按"意思"而非字面搜索商品、文档、问答
推荐系统	“喜欢 X 的人也喜欢……”，找相似向量
文本分类	用最近邻的类别投票归类
去重 / 异常检测	找近重复内容；找离正常模式很远的异常点

二、Redis 向量检索的发展与整体架构

2.1 演进时间线

RediSearch 2.4（2021）：首次引入向量相似度搜索。
Redis Stack：把 RediSearch、RedisJSON 等模块打包。
Redis 8（2025）：查询引擎（Query Engine）成为 Redis 标准组件，开源可用，无需单独装模块；同时新增原生数据类型 Vector Sets。

2.2 Redis 提供两套向量能力

能力别名存储特点

Query Engine	原 RediSearch	HASH / JSON	功能最全：多字段、全文、标签、地理、混合查询
Vector Sets（VSET）	原生数据类型	自带	类似有序集合，内置量化（Q8/BIN），轻量省内存

#mermaid-svg-dLJDcEbJMY2Or5Dl{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-dLJDcEbJMY2Or5Dl .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-dLJDcEbJMY2Or5Dl .error-icon{fill:#552222;}#mermaid-svg-dLJDcEbJMY2Or5Dl .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-dLJDcEbJMY2Or5Dl .marker{fill:#333333;stroke:#333333;}#mermaid-svg-dLJDcEbJMY2Or5Dl .marker.cross{stroke:#333333;}#mermaid-svg-dLJDcEbJMY2Or5Dl svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-dLJDcEbJMY2Or5Dl p{margin:0;}#mermaid-svg-dLJDcEbJMY2Or5Dl .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-dLJDcEbJMY2Or5Dl .cluster-label text{fill:#333;}#mermaid-svg-dLJDcEbJMY2Or5Dl .cluster-label span{color:#333;}#mermaid-svg-dLJDcEbJMY2Or5Dl .cluster-label span p{background-color:transparent;}#mermaid-svg-dLJDcEbJMY2Or5Dl .label text,#mermaid-svg-dLJDcEbJMY2Or5Dl span{fill:#333;color:#333;}#mermaid-svg-dLJDcEbJMY2Or5Dl .node rect,#mermaid-svg-dLJDcEbJMY2Or5Dl .node circle,#mermaid-svg-dLJDcEbJMY2Or5Dl .node ellipse,#mermaid-svg-dLJDcEbJMY2Or5Dl .node polygon,#mermaid-svg-dLJDcEbJMY2Or5Dl .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-dLJDcEbJMY2Or5Dl .rough-node .label text,#mermaid-svg-dLJDcEbJMY2Or5Dl .node .label text,#mermaid-svg-dLJDcEbJMY2Or5Dl .image-shape .label,#mermaid-svg-dLJDcEbJMY2Or5Dl .icon-shape .label{text-anchor:middle;}#mermaid-svg-dLJDcEbJMY2Or5Dl .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-dLJDcEbJMY2Or5Dl .rough-node .label,#mermaid-svg-dLJDcEbJMY2Or5Dl .node .label,#mermaid-svg-dLJDcEbJMY2Or5Dl .image-shape .label,#mermaid-svg-dLJDcEbJMY2Or5Dl .icon-shape .label{text-align:center;}#mermaid-svg-dLJDcEbJMY2Or5Dl .node.clickable{cursor:pointer;}#mermaid-svg-dLJDcEbJMY2Or5Dl .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-dLJDcEbJMY2Or5Dl .arrowheadPath{fill:#333333;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-dLJDcEbJMY2Or5Dl .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-dLJDcEbJMY2Or5Dl .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-dLJDcEbJMY2Or5Dl .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-dLJDcEbJMY2Or5Dl .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-dLJDcEbJMY2Or5Dl .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-dLJDcEbJMY2Or5Dl .cluster text{fill:#333;}#mermaid-svg-dLJDcEbJMY2Or5Dl .cluster span{color:#333;}#mermaid-svg-dLJDcEbJMY2Or5Dl 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-dLJDcEbJMY2Or5Dl .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-dLJDcEbJMY2Or5Dl rect.text{fill:none;stroke-width:0;}#mermaid-svg-dLJDcEbJMY2Or5Dl .icon-shape,#mermaid-svg-dLJDcEbJMY2Or5Dl .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-dLJDcEbJMY2Or5Dl .icon-shape p,#mermaid-svg-dLJDcEbJMY2Or5Dl .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-dLJDcEbJMY2Or5Dl .icon-shape .label rect,#mermaid-svg-dLJDcEbJMY2Or5Dl .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-dLJDcEbJMY2Or5Dl .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-dLJDcEbJMY2Or5Dl .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-dLJDcEbJMY2Or5Dl :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

应用

Query Engine（原 RediSearch）

Vector Sets 原生数据类型

HASH 存储：向量为原始字节

JSON 存储：向量为浮点数组

VSET：内置 HNSW 与量化

怎么选：需要把向量和文本、标签、数值等多字段做复杂混合查询 → Query Engine；只做"给向量找相似"、追求极致省内存和简单 → Vector Sets。

三、核心原理：Embedding 与相似度

3.1 向量数据类型

类型说明

FLOAT32	最常用，单精度，精度与内存平衡
FLOAT64	双精度，更准但更占内存
FLOAT16 / BFLOAT16	半精度，省内存（需查询引擎 v2.10+）

3.2 三种距离度量

设两个向量 u、v：

度量含义公式适用

L2	欧几里得距离（直线距离）	根号下各维度差的平方和	通用、空间距离直观
IP	内积	1 减去 u·v	向量已归一化时
COSINE	余弦距离（看方向夹角）	1 减去余弦相似度	文本语义最常用，不受长度影响

#mermaid-svg-XYkiUB83mihoSn0a{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-XYkiUB83mihoSn0a .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-XYkiUB83mihoSn0a .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-XYkiUB83mihoSn0a .error-icon{fill:#552222;}#mermaid-svg-XYkiUB83mihoSn0a .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-XYkiUB83mihoSn0a .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-XYkiUB83mihoSn0a .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-XYkiUB83mihoSn0a .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-XYkiUB83mihoSn0a .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-XYkiUB83mihoSn0a .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-XYkiUB83mihoSn0a .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-XYkiUB83mihoSn0a .marker{fill:#333333;stroke:#333333;}#mermaid-svg-XYkiUB83mihoSn0a .marker.cross{stroke:#333333;}#mermaid-svg-XYkiUB83mihoSn0a svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-XYkiUB83mihoSn0a p{margin:0;}#mermaid-svg-XYkiUB83mihoSn0a .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-XYkiUB83mihoSn0a .cluster-label text{fill:#333;}#mermaid-svg-XYkiUB83mihoSn0a .cluster-label span{color:#333;}#mermaid-svg-XYkiUB83mihoSn0a .cluster-label span p{background-color:transparent;}#mermaid-svg-XYkiUB83mihoSn0a .label text,#mermaid-svg-XYkiUB83mihoSn0a span{fill:#333;color:#333;}#mermaid-svg-XYkiUB83mihoSn0a .node rect,#mermaid-svg-XYkiUB83mihoSn0a .node circle,#mermaid-svg-XYkiUB83mihoSn0a .node ellipse,#mermaid-svg-XYkiUB83mihoSn0a .node polygon,#mermaid-svg-XYkiUB83mihoSn0a .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-XYkiUB83mihoSn0a .rough-node .label text,#mermaid-svg-XYkiUB83mihoSn0a .node .label text,#mermaid-svg-XYkiUB83mihoSn0a .image-shape .label,#mermaid-svg-XYkiUB83mihoSn0a .icon-shape .label{text-anchor:middle;}#mermaid-svg-XYkiUB83mihoSn0a .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-XYkiUB83mihoSn0a .rough-node .label,#mermaid-svg-XYkiUB83mihoSn0a .node .label,#mermaid-svg-XYkiUB83mihoSn0a .image-shape .label,#mermaid-svg-XYkiUB83mihoSn0a .icon-shape .label{text-align:center;}#mermaid-svg-XYkiUB83mihoSn0a .node.clickable{cursor:pointer;}#mermaid-svg-XYkiUB83mihoSn0a .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-XYkiUB83mihoSn0a .arrowheadPath{fill:#333333;}#mermaid-svg-XYkiUB83mihoSn0a .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-XYkiUB83mihoSn0a .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-XYkiUB83mihoSn0a .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-XYkiUB83mihoSn0a .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-XYkiUB83mihoSn0a .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-XYkiUB83mihoSn0a .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-XYkiUB83mihoSn0a .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-XYkiUB83mihoSn0a .cluster text{fill:#333;}#mermaid-svg-XYkiUB83mihoSn0a .cluster span{color:#333;}#mermaid-svg-XYkiUB83mihoSn0a 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-XYkiUB83mihoSn0a .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-XYkiUB83mihoSn0a rect.text{fill:none;stroke-width:0;}#mermaid-svg-XYkiUB83mihoSn0a .icon-shape,#mermaid-svg-XYkiUB83mihoSn0a .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-XYkiUB83mihoSn0a .icon-shape p,#mermaid-svg-XYkiUB83mihoSn0a .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-XYkiUB83mihoSn0a .icon-shape .label rect,#mermaid-svg-XYkiUB83mihoSn0a .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-XYkiUB83mihoSn0a .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-XYkiUB83mihoSn0a .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-XYkiUB83mihoSn0a :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

两个向量

选择度量

L2：看直线距离

IP：看内积

COSINE：看方向夹角，文本首选

关键理解：

距离值越小越相似。
余弦距离上限为 2，L2 无上限。
文本 Embedding 通常已经归一化，此时 COSINE 与 IP 等价，绝大多数 RAG 场景选 COSINE。

四、索引算法原理：FLAT 与 HNSW

4.1 精确 vs 近似：KNN 与 ANN

KNN（精确最近邻）：拿查询向量和所有向量逐一比较，保证找到真值，但数据量大时极慢。
ANN（近似最近邻）：用特殊索引结构快速逼近，牺牲极小准确率（召回率通常 95% 到 99%），换数量级的提速。

4.2 FLAT 索引（暴力精确）

查询时把查询向量与每一个已存向量比较、排序，返回前 K 个。结果 100% 准确，但耗时随数据量线性增长。

#mermaid-svg-QR1oZUTrJsKFnvAm{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-QR1oZUTrJsKFnvAm .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-QR1oZUTrJsKFnvAm .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-QR1oZUTrJsKFnvAm .error-icon{fill:#552222;}#mermaid-svg-QR1oZUTrJsKFnvAm .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-QR1oZUTrJsKFnvAm .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-QR1oZUTrJsKFnvAm .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-QR1oZUTrJsKFnvAm .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-QR1oZUTrJsKFnvAm .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-QR1oZUTrJsKFnvAm .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-QR1oZUTrJsKFnvAm .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-QR1oZUTrJsKFnvAm .marker{fill:#333333;stroke:#333333;}#mermaid-svg-QR1oZUTrJsKFnvAm .marker.cross{stroke:#333333;}#mermaid-svg-QR1oZUTrJsKFnvAm svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-QR1oZUTrJsKFnvAm p{margin:0;}#mermaid-svg-QR1oZUTrJsKFnvAm .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-QR1oZUTrJsKFnvAm .cluster-label text{fill:#333;}#mermaid-svg-QR1oZUTrJsKFnvAm .cluster-label span{color:#333;}#mermaid-svg-QR1oZUTrJsKFnvAm .cluster-label span p{background-color:transparent;}#mermaid-svg-QR1oZUTrJsKFnvAm .label text,#mermaid-svg-QR1oZUTrJsKFnvAm span{fill:#333;color:#333;}#mermaid-svg-QR1oZUTrJsKFnvAm .node rect,#mermaid-svg-QR1oZUTrJsKFnvAm .node circle,#mermaid-svg-QR1oZUTrJsKFnvAm .node ellipse,#mermaid-svg-QR1oZUTrJsKFnvAm .node polygon,#mermaid-svg-QR1oZUTrJsKFnvAm .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-QR1oZUTrJsKFnvAm .rough-node .label text,#mermaid-svg-QR1oZUTrJsKFnvAm .node .label text,#mermaid-svg-QR1oZUTrJsKFnvAm .image-shape .label,#mermaid-svg-QR1oZUTrJsKFnvAm .icon-shape .label{text-anchor:middle;}#mermaid-svg-QR1oZUTrJsKFnvAm .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-QR1oZUTrJsKFnvAm .rough-node .label,#mermaid-svg-QR1oZUTrJsKFnvAm .node .label,#mermaid-svg-QR1oZUTrJsKFnvAm .image-shape .label,#mermaid-svg-QR1oZUTrJsKFnvAm .icon-shape .label{text-align:center;}#mermaid-svg-QR1oZUTrJsKFnvAm .node.clickable{cursor:pointer;}#mermaid-svg-QR1oZUTrJsKFnvAm .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-QR1oZUTrJsKFnvAm .arrowheadPath{fill:#333333;}#mermaid-svg-QR1oZUTrJsKFnvAm .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-QR1oZUTrJsKFnvAm .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-QR1oZUTrJsKFnvAm .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-QR1oZUTrJsKFnvAm .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-QR1oZUTrJsKFnvAm .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-QR1oZUTrJsKFnvAm .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-QR1oZUTrJsKFnvAm .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-QR1oZUTrJsKFnvAm .cluster text{fill:#333;}#mermaid-svg-QR1oZUTrJsKFnvAm .cluster span{color:#333;}#mermaid-svg-QR1oZUTrJsKFnvAm 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-QR1oZUTrJsKFnvAm .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-QR1oZUTrJsKFnvAm rect.text{fill:none;stroke-width:0;}#mermaid-svg-QR1oZUTrJsKFnvAm .icon-shape,#mermaid-svg-QR1oZUTrJsKFnvAm .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-QR1oZUTrJsKFnvAm .icon-shape p,#mermaid-svg-QR1oZUTrJsKFnvAm .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-QR1oZUTrJsKFnvAm .icon-shape .label rect,#mermaid-svg-QR1oZUTrJsKFnvAm .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-QR1oZUTrJsKFnvAm .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-QR1oZUTrJsKFnvAm .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-QR1oZUTrJsKFnvAm :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

查询向量

与向量1比较

与向量2比较

与向量N比较……

全部排序

返回最近K个（精确）

适用：小数据集（约 10 万以下）、开发调试、对准确率要求极高。
1 万向量约 10ms，100 万可能到 1000ms。

4.3 HNSW 索引（近似，工业界首选）

HNSW（Hierarchical Navigable Small World，分层可导航小世界）由 Malkov 和 Yashunin 在 2016 年论文提出，是当前综合性能最优、应用最广的向量索引算法。

4.3.1 核心思想：高速公路 + 城市道路

构建一个多层图：

第 0 层（最底层）：包含全部节点，连接最密集（“城市道路”）。
越往上：节点越少（随机采样的子集），连接越稀疏、跨度越大（“高速公路”），用于快速定位大致区域。

#mermaid-svg-lwzIWnrNckwWDz83{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-lwzIWnrNckwWDz83 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-lwzIWnrNckwWDz83 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-lwzIWnrNckwWDz83 .error-icon{fill:#552222;}#mermaid-svg-lwzIWnrNckwWDz83 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-lwzIWnrNckwWDz83 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-lwzIWnrNckwWDz83 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-lwzIWnrNckwWDz83 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-lwzIWnrNckwWDz83 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-lwzIWnrNckwWDz83 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-lwzIWnrNckwWDz83 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-lwzIWnrNckwWDz83 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-lwzIWnrNckwWDz83 .marker.cross{stroke:#333333;}#mermaid-svg-lwzIWnrNckwWDz83 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-lwzIWnrNckwWDz83 p{margin:0;}#mermaid-svg-lwzIWnrNckwWDz83 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-lwzIWnrNckwWDz83 .cluster-label text{fill:#333;}#mermaid-svg-lwzIWnrNckwWDz83 .cluster-label span{color:#333;}#mermaid-svg-lwzIWnrNckwWDz83 .cluster-label span p{background-color:transparent;}#mermaid-svg-lwzIWnrNckwWDz83 .label text,#mermaid-svg-lwzIWnrNckwWDz83 span{fill:#333;color:#333;}#mermaid-svg-lwzIWnrNckwWDz83 .node rect,#mermaid-svg-lwzIWnrNckwWDz83 .node circle,#mermaid-svg-lwzIWnrNckwWDz83 .node ellipse,#mermaid-svg-lwzIWnrNckwWDz83 .node polygon,#mermaid-svg-lwzIWnrNckwWDz83 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-lwzIWnrNckwWDz83 .rough-node .label text,#mermaid-svg-lwzIWnrNckwWDz83 .node .label text,#mermaid-svg-lwzIWnrNckwWDz83 .image-shape .label,#mermaid-svg-lwzIWnrNckwWDz83 .icon-shape .label{text-anchor:middle;}#mermaid-svg-lwzIWnrNckwWDz83 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-lwzIWnrNckwWDz83 .rough-node .label,#mermaid-svg-lwzIWnrNckwWDz83 .node .label,#mermaid-svg-lwzIWnrNckwWDz83 .image-shape .label,#mermaid-svg-lwzIWnrNckwWDz83 .icon-shape .label{text-align:center;}#mermaid-svg-lwzIWnrNckwWDz83 .node.clickable{cursor:pointer;}#mermaid-svg-lwzIWnrNckwWDz83 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-lwzIWnrNckwWDz83 .arrowheadPath{fill:#333333;}#mermaid-svg-lwzIWnrNckwWDz83 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-lwzIWnrNckwWDz83 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-lwzIWnrNckwWDz83 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-lwzIWnrNckwWDz83 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-lwzIWnrNckwWDz83 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-lwzIWnrNckwWDz83 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-lwzIWnrNckwWDz83 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-lwzIWnrNckwWDz83 .cluster text{fill:#333;}#mermaid-svg-lwzIWnrNckwWDz83 .cluster span{color:#333;}#mermaid-svg-lwzIWnrNckwWDz83 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-lwzIWnrNckwWDz83 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-lwzIWnrNckwWDz83 rect.text{fill:none;stroke-width:0;}#mermaid-svg-lwzIWnrNckwWDz83 .icon-shape,#mermaid-svg-lwzIWnrNckwWDz83 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-lwzIWnrNckwWDz83 .icon-shape p,#mermaid-svg-lwzIWnrNckwWDz83 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-lwzIWnrNckwWDz83 .icon-shape .label rect,#mermaid-svg-lwzIWnrNckwWDz83 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-lwzIWnrNckwWDz83 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-lwzIWnrNckwWDz83 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-lwzIWnrNckwWDz83 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

第0层：全量节点

节点

第1层：子集节点

节点

第2层：最稀疏，高速入口

入口节点

4.3.2 查询原理：从顶层贪心、逐层下降

#mermaid-svg-1QALPOZ91pcYbNR6{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-1QALPOZ91pcYbNR6 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-1QALPOZ91pcYbNR6 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-1QALPOZ91pcYbNR6 .error-icon{fill:#552222;}#mermaid-svg-1QALPOZ91pcYbNR6 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-1QALPOZ91pcYbNR6 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-1QALPOZ91pcYbNR6 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-1QALPOZ91pcYbNR6 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-1QALPOZ91pcYbNR6 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-1QALPOZ91pcYbNR6 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-1QALPOZ91pcYbNR6 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-1QALPOZ91pcYbNR6 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-1QALPOZ91pcYbNR6 .marker.cross{stroke:#333333;}#mermaid-svg-1QALPOZ91pcYbNR6 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-1QALPOZ91pcYbNR6 p{margin:0;}#mermaid-svg-1QALPOZ91pcYbNR6 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-1QALPOZ91pcYbNR6 .cluster-label text{fill:#333;}#mermaid-svg-1QALPOZ91pcYbNR6 .cluster-label span{color:#333;}#mermaid-svg-1QALPOZ91pcYbNR6 .cluster-label span p{background-color:transparent;}#mermaid-svg-1QALPOZ91pcYbNR6 .label text,#mermaid-svg-1QALPOZ91pcYbNR6 span{fill:#333;color:#333;}#mermaid-svg-1QALPOZ91pcYbNR6 .node rect,#mermaid-svg-1QALPOZ91pcYbNR6 .node circle,#mermaid-svg-1QALPOZ91pcYbNR6 .node ellipse,#mermaid-svg-1QALPOZ91pcYbNR6 .node polygon,#mermaid-svg-1QALPOZ91pcYbNR6 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-1QALPOZ91pcYbNR6 .rough-node .label text,#mermaid-svg-1QALPOZ91pcYbNR6 .node .label text,#mermaid-svg-1QALPOZ91pcYbNR6 .image-shape .label,#mermaid-svg-1QALPOZ91pcYbNR6 .icon-shape .label{text-anchor:middle;}#mermaid-svg-1QALPOZ91pcYbNR6 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-1QALPOZ91pcYbNR6 .rough-node .label,#mermaid-svg-1QALPOZ91pcYbNR6 .node .label,#mermaid-svg-1QALPOZ91pcYbNR6 .image-shape .label,#mermaid-svg-1QALPOZ91pcYbNR6 .icon-shape .label{text-align:center;}#mermaid-svg-1QALPOZ91pcYbNR6 .node.clickable{cursor:pointer;}#mermaid-svg-1QALPOZ91pcYbNR6 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-1QALPOZ91pcYbNR6 .arrowheadPath{fill:#333333;}#mermaid-svg-1QALPOZ91pcYbNR6 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-1QALPOZ91pcYbNR6 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-1QALPOZ91pcYbNR6 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-1QALPOZ91pcYbNR6 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-1QALPOZ91pcYbNR6 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-1QALPOZ91pcYbNR6 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-1QALPOZ91pcYbNR6 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-1QALPOZ91pcYbNR6 .cluster text{fill:#333;}#mermaid-svg-1QALPOZ91pcYbNR6 .cluster span{color:#333;}#mermaid-svg-1QALPOZ91pcYbNR6 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-1QALPOZ91pcYbNR6 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-1QALPOZ91pcYbNR6 rect.text{fill:none;stroke-width:0;}#mermaid-svg-1QALPOZ91pcYbNR6 .icon-shape,#mermaid-svg-1QALPOZ91pcYbNR6 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-1QALPOZ91pcYbNR6 .icon-shape p,#mermaid-svg-1QALPOZ91pcYbNR6 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-1QALPOZ91pcYbNR6 .icon-shape .label rect,#mermaid-svg-1QALPOZ91pcYbNR6 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-1QALPOZ91pcYbNR6 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-1QALPOZ91pcYbNR6 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-1QALPOZ91pcYbNR6 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

是

否

从最顶层入口节点开始

在本层贪心走向离查询向量最近的邻居

找到本层局部最近点

还有更底层吗？

下降到下一层，以该点为新起点

返回第0层找到的最近K个节点

就像导航：先走高速公路快速接近目标城市，再逐层下到城市道路、街道，最后精确找到目的地。这样把搜索复杂度降到约 O(log N)。

4.3.3 关键参数

参数含义默认调大的影响

M	每个节点每层最大连接数（第0层为 2M）	16	召回更高、但内存和构建时间增加
EF_CONSTRUCTION	建图时考虑的邻居数	200	图质量更高、建图更慢
EF_RUNTIME	查询时保留的候选数	10	召回更高、查询更慢

调优口诀：召回不够先调大 EF_RUNTIME（无需重建索引）；还不够再调大 M 并重建。

4.3.4 召回率

HNSW 通常能达到 95% 到 99% 召回，即在 100 个真实最近邻中能找到 95 到 99 个，而速度比 FLAT 快几个数量级。

五、实战一：Query Engine（FT 命令）

5.1 创建索引

FT.CREATE documents
ON HASH
PREFIX 1 docs:
SCHEMA
title TEXT
category TAG
year NUMERIC
embedding VECTOR HNSW 10
TYPE FLOAT32
DIM 1536
DISTANCE_METRIC COSINE
M 16
EF_CONSTRUCTION 200

5.2 写入向量（HASH）

HSET docs:01 embedding <vector_bytes> category sports year 2021

5.3 KNN 查询

FT.SEARCH documents
"*=>[KNN 10 @embedding $BLOB AS vector_distance]"
PARAMS 2 BLOB "\\x12\\xa9\\xf5…"
SORTBY vector_distance
LIMIT 0 10
DIALECT 2

5.4 混合查询（向量 + 元数据过滤）

FT.SEARCH documents
"(@title:Dune @year:[2020 2022])=>[KNN 10 @embedding $BLOB AS vector_distance]"
PARAMS 2 BLOB "\\x12\\xa9\\xf5…"
SORTBY vector_distance
LIMIT 0 10
DIALECT 2

过滤的两种内部模式：

#mermaid-svg-jcK3TXqXwLmoeIxp{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-jcK3TXqXwLmoeIxp .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-jcK3TXqXwLmoeIxp .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-jcK3TXqXwLmoeIxp .error-icon{fill:#552222;}#mermaid-svg-jcK3TXqXwLmoeIxp .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-jcK3TXqXwLmoeIxp .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-jcK3TXqXwLmoeIxp .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-jcK3TXqXwLmoeIxp .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-jcK3TXqXwLmoeIxp .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-jcK3TXqXwLmoeIxp .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-jcK3TXqXwLmoeIxp .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-jcK3TXqXwLmoeIxp .marker{fill:#333333;stroke:#333333;}#mermaid-svg-jcK3TXqXwLmoeIxp .marker.cross{stroke:#333333;}#mermaid-svg-jcK3TXqXwLmoeIxp svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-jcK3TXqXwLmoeIxp p{margin:0;}#mermaid-svg-jcK3TXqXwLmoeIxp .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-jcK3TXqXwLmoeIxp .cluster-label text{fill:#333;}#mermaid-svg-jcK3TXqXwLmoeIxp .cluster-label span{color:#333;}#mermaid-svg-jcK3TXqXwLmoeIxp .cluster-label span p{background-color:transparent;}#mermaid-svg-jcK3TXqXwLmoeIxp .label text,#mermaid-svg-jcK3TXqXwLmoeIxp span{fill:#333;color:#333;}#mermaid-svg-jcK3TXqXwLmoeIxp .node rect,#mermaid-svg-jcK3TXqXwLmoeIxp .node circle,#mermaid-svg-jcK3TXqXwLmoeIxp .node ellipse,#mermaid-svg-jcK3TXqXwLmoeIxp .node polygon,#mermaid-svg-jcK3TXqXwLmoeIxp .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-jcK3TXqXwLmoeIxp .rough-node .label text,#mermaid-svg-jcK3TXqXwLmoeIxp .node .label text,#mermaid-svg-jcK3TXqXwLmoeIxp .image-shape .label,#mermaid-svg-jcK3TXqXwLmoeIxp .icon-shape .label{text-anchor:middle;}#mermaid-svg-jcK3TXqXwLmoeIxp .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-jcK3TXqXwLmoeIxp .rough-node .label,#mermaid-svg-jcK3TXqXwLmoeIxp .node .label,#mermaid-svg-jcK3TXqXwLmoeIxp .image-shape .label,#mermaid-svg-jcK3TXqXwLmoeIxp .icon-shape .label{text-align:center;}#mermaid-svg-jcK3TXqXwLmoeIxp .node.clickable{cursor:pointer;}#mermaid-svg-jcK3TXqXwLmoeIxp .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-jcK3TXqXwLmoeIxp .arrowheadPath{fill:#333333;}#mermaid-svg-jcK3TXqXwLmoeIxp .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-jcK3TXqXwLmoeIxp .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-jcK3TXqXwLmoeIxp .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-jcK3TXqXwLmoeIxp .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-jcK3TXqXwLmoeIxp .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-jcK3TXqXwLmoeIxp .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-jcK3TXqXwLmoeIxp .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-jcK3TXqXwLmoeIxp .cluster text{fill:#333;}#mermaid-svg-jcK3TXqXwLmoeIxp .cluster span{color:#333;}#mermaid-svg-jcK3TXqXwLmoeIxp 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-jcK3TXqXwLmoeIxp .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-jcK3TXqXwLmoeIxp rect.text{fill:none;stroke-width:0;}#mermaid-svg-jcK3TXqXwLmoeIxp .icon-shape,#mermaid-svg-jcK3TXqXwLmoeIxp .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-jcK3TXqXwLmoeIxp .icon-shape p,#mermaid-svg-jcK3TXqXwLmoeIxp .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-jcK3TXqXwLmoeIxp .icon-shape .label rect,#mermaid-svg-jcK3TXqXwLmoeIxp .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-jcK3TXqXwLmoeIxp .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-jcK3TXqXwLmoeIxp .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-jcK3TXqXwLmoeIxp :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

带过滤的向量查询

BATCHES 批量模式：分页取最近邻，边过滤边收集

ADHOC_BF 临时暴力：过滤后文档少时直接精算，结果准确

5.5 Python 完整示例（redis-py）

import numpy as np
from redis import Redis
from redis.commands.search.field import VectorField, TagField
from redis.commands.search.indexDefinition import IndexDefinition, IndexType
from redis.commands.search.query import Query

r = Redis(host="localhost", port=6379, decode_responses=False)

# 1. 创建 HNSW 索引
schema = [
TagField("category"),
VectorField("embedding", "HNSW", {
"TYPE": "FLOAT32",
"DIM": 1536,
"DISTANCE_METRIC": "COSINE",
"M": 16,
"EF_CONSTRUCTION": 200,
}),
]
r.ft("documents").create_index(
schema,
definition=IndexDefinition(prefix=["docs:"], index_type=IndexType.HASH),
)

# 2. 写入一条向量（FLOAT32 转字节）
vec = np.array([0.12, –0.34, 0.56], dtype=np.float32) # 实际为 1536 维
r.hset("docs:01", mapping={"embedding": vec.tobytes(), "category": "sports"})

# 3. KNN 向量检索（运行时调大 EF_RUNTIME 提升召回）
k = 10
query = (
Query("*=>[KNN $K @embedding $B]=>{$EF_RUNTIME:$EF}")
.sort_by("embedding")
.paging(0, k)
.dialect(2)
)
result = r.ft("documents").search(
query,
query_params={"B": vec.tobytes(), "K": k, "EF": 200},
)
for doc in result.docs:
print(doc.id, doc.embedding_score)

5.6 Java 示例（Spring AI，概念）

Spring AI 提供 RedisVectorStore，底层自动完成索引创建、写入与 KNN 检索：

/**
* 配置 Redis 向量存储（Spring AI）
* 底层自动调用 FT.CREATE 建立 HNSW 索引
*/
@Bean
public VectorStore vectorStore(JedisConnectionFactory factory, EmbeddingModel embeddingModel) {
RedisVectorStoreConfig config = RedisVectorStoreConfig.builder()
.indexName("documents")
.prefix("docs:")
.indexAlgorithm(IndexAlgorithm.HNSW)
.build();
return new RedisVectorStore(config, factory, embeddingModel, true);
}

// 业务中直接检索，无需手写 FT 命令
List<Document> hits = vectorStore.similaritySearch(
SearchRequest.query("如何提升系统性能").withTopK(5));

提示：不同 Spring AI 版本的配置类略有差异，以官方文档为准；需要精细控制时可直接用 Jedis 的 FT_SEARCH 原生命令。

六、实战二：Vector Sets（VSET 命令）

6.1 添加与检索

# 添加向量（默认 Q8 量化，省4倍内存）
VADD movies VALUES 3 0.1 0.5 0.9 "inception" SETATTR '{"year":2010,"genre":"scifi"}'

# 找相似（按向量）
VSIM movies VALUES 3 0.1 0.5 0.9 COUNT 10 WITHSCORES

# 按已有元素找相似
VSIM movies ELE "inception" COUNT 10 WITHSCORES

# 带属性过滤
VSIM movies VALUES 3 0.1 0.5 0.9 FILTER '.year >= 2000 and .genre == "scifi"' COUNT 10

6.2 常用命令

命令作用

VADD	添加元素和向量
VSIM	找相似元素
VREM	删除（真删除，内存回收）
VCARD	元素数量
VDIM	向量维度
VEMB	取某元素向量
VINFO	索引元信息
VSETATTR / VGETATTR	设置 / 读取 JSON 属性

6.3 量化方式

量化内存速度召回

NOQUANT（FP32）	4 字节/维	基准	最好
Q8（默认）	1 字节/维	约 2 倍	约 96%
BIN（二值）	1 比特/维	约 4 倍	约 80%

6.4 Python 示例

import redis
r = redis.Redis()

# 添加（默认 Q8 量化）
r.execute_command("VADD", "movies", "VALUES", 3, 0.1, 0.5, 0.9,
"inception", "SETATTR", '{"year":2010,"genre":"scifi"}')

# 检索
res = r.execute_command("VSIM", "movies", "VALUES", 3, 0.1, 0.5, 0.9,
"COUNT", 10, "WITHSCORES")
print(res)

七、性能、内存与调优

7.1 内存估算

以 300 维、默认 Q8（Vector Sets）为例：

向量存储：约 1 字节/维。
图开销：约每个元素 M×2.5 个指针。
合计：约 1KB / 元素。

Query Engine 用 FLOAT32 时约 4 字节/维，再加 HNSW 图连接。

7.2 调优清单

#mermaid-svg-cLmUWtfX0zzRFiPC{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-cLmUWtfX0zzRFiPC .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-cLmUWtfX0zzRFiPC .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-cLmUWtfX0zzRFiPC .error-icon{fill:#552222;}#mermaid-svg-cLmUWtfX0zzRFiPC .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-cLmUWtfX0zzRFiPC .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-cLmUWtfX0zzRFiPC .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-cLmUWtfX0zzRFiPC .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-cLmUWtfX0zzRFiPC .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-cLmUWtfX0zzRFiPC .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-cLmUWtfX0zzRFiPC .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-cLmUWtfX0zzRFiPC .marker{fill:#333333;stroke:#333333;}#mermaid-svg-cLmUWtfX0zzRFiPC .marker.cross{stroke:#333333;}#mermaid-svg-cLmUWtfX0zzRFiPC svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-cLmUWtfX0zzRFiPC p{margin:0;}#mermaid-svg-cLmUWtfX0zzRFiPC .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-cLmUWtfX0zzRFiPC .cluster-label text{fill:#333;}#mermaid-svg-cLmUWtfX0zzRFiPC .cluster-label span{color:#333;}#mermaid-svg-cLmUWtfX0zzRFiPC .cluster-label span p{background-color:transparent;}#mermaid-svg-cLmUWtfX0zzRFiPC .label text,#mermaid-svg-cLmUWtfX0zzRFiPC span{fill:#333;color:#333;}#mermaid-svg-cLmUWtfX0zzRFiPC .node rect,#mermaid-svg-cLmUWtfX0zzRFiPC .node circle,#mermaid-svg-cLmUWtfX0zzRFiPC .node ellipse,#mermaid-svg-cLmUWtfX0zzRFiPC .node polygon,#mermaid-svg-cLmUWtfX0zzRFiPC .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-cLmUWtfX0zzRFiPC .rough-node .label text,#mermaid-svg-cLmUWtfX0zzRFiPC .node .label text,#mermaid-svg-cLmUWtfX0zzRFiPC .image-shape .label,#mermaid-svg-cLmUWtfX0zzRFiPC .icon-shape .label{text-anchor:middle;}#mermaid-svg-cLmUWtfX0zzRFiPC .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-cLmUWtfX0zzRFiPC .rough-node .label,#mermaid-svg-cLmUWtfX0zzRFiPC .node .label,#mermaid-svg-cLmUWtfX0zzRFiPC .image-shape .label,#mermaid-svg-cLmUWtfX0zzRFiPC .icon-shape .label{text-align:center;}#mermaid-svg-cLmUWtfX0zzRFiPC .node.clickable{cursor:pointer;}#mermaid-svg-cLmUWtfX0zzRFiPC .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-cLmUWtfX0zzRFiPC .arrowheadPath{fill:#333333;}#mermaid-svg-cLmUWtfX0zzRFiPC .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-cLmUWtfX0zzRFiPC .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-cLmUWtfX0zzRFiPC .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-cLmUWtfX0zzRFiPC .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-cLmUWtfX0zzRFiPC .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-cLmUWtfX0zzRFiPC .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-cLmUWtfX0zzRFiPC .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-cLmUWtfX0zzRFiPC .cluster text{fill:#333;}#mermaid-svg-cLmUWtfX0zzRFiPC .cluster span{color:#333;}#mermaid-svg-cLmUWtfX0zzRFiPC 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-cLmUWtfX0zzRFiPC .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-cLmUWtfX0zzRFiPC rect.text{fill:none;stroke-width:0;}#mermaid-svg-cLmUWtfX0zzRFiPC .icon-shape,#mermaid-svg-cLmUWtfX0zzRFiPC .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-cLmUWtfX0zzRFiPC .icon-shape p,#mermaid-svg-cLmUWtfX0zzRFiPC .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-cLmUWtfX0zzRFiPC .icon-shape .label rect,#mermaid-svg-cLmUWtfX0zzRFiPC .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-cLmUWtfX0zzRFiPC .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-cLmUWtfX0zzRFiPC .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-cLmUWtfX0zzRFiPC :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

是

否

是

否

性能/召回问题

召回率不足？

调大 EF_RUNTIME（不重建）

仍不足？

调大 M 并重建索引

内存吃紧？

用 Q8 或 BIN 量化 / 降维

完成

召回优先：先调 EF_RUNTIME，再调 M。

内存优先：用 Q8（影响小）或 BIN（配合重排）；高维可 REDUCE 降维。

吞吐量：官方数据 Vector Sets 约 5 万次 VSIM/秒（300 万、300 维）。

召回调试：用 VSIM … TRUTH 取精确结果，与近似结果对比算召回。

7.3 水平扩展（分片）

#mermaid-svg-DXpGr3p3RBWg8aJ0{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-DXpGr3p3RBWg8aJ0 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .error-icon{fill:#552222;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .marker.cross{stroke:#333333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-DXpGr3p3RBWg8aJ0 p{margin:0;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .cluster-label text{fill:#333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .cluster-label span{color:#333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .cluster-label span p{background-color:transparent;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .label text,#mermaid-svg-DXpGr3p3RBWg8aJ0 span{fill:#333;color:#333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .node rect,#mermaid-svg-DXpGr3p3RBWg8aJ0 .node circle,#mermaid-svg-DXpGr3p3RBWg8aJ0 .node ellipse,#mermaid-svg-DXpGr3p3RBWg8aJ0 .node polygon,#mermaid-svg-DXpGr3p3RBWg8aJ0 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .rough-node .label text,#mermaid-svg-DXpGr3p3RBWg8aJ0 .node .label text,#mermaid-svg-DXpGr3p3RBWg8aJ0 .image-shape .label,#mermaid-svg-DXpGr3p3RBWg8aJ0 .icon-shape .label{text-anchor:middle;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .rough-node .label,#mermaid-svg-DXpGr3p3RBWg8aJ0 .node .label,#mermaid-svg-DXpGr3p3RBWg8aJ0 .image-shape .label,#mermaid-svg-DXpGr3p3RBWg8aJ0 .icon-shape .label{text-align:center;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .node.clickable{cursor:pointer;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .arrowheadPath{fill:#333333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-DXpGr3p3RBWg8aJ0 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-DXpGr3p3RBWg8aJ0 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-DXpGr3p3RBWg8aJ0 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .cluster text{fill:#333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .cluster span{color:#333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 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-DXpGr3p3RBWg8aJ0 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-DXpGr3p3RBWg8aJ0 rect.text{fill:none;stroke-width:0;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .icon-shape,#mermaid-svg-DXpGr3p3RBWg8aJ0 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .icon-shape p,#mermaid-svg-DXpGr3p3RBWg8aJ0 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .icon-shape .label rect,#mermaid-svg-DXpGr3p3RBWg8aJ0 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-DXpGr3p3RBWg8aJ0 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-DXpGr3p3RBWg8aJ0 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-DXpGr3p3RBWg8aJ0 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

查询

分片0

分片1

分片2

并行查询，按分数合并取TopK

按元素哈希分片到多个 Redis 实例，并行查询、客户端合并，写入线性扩展。

八、与其他向量数据库对比及选型

8.1 主流产品对比

产品定位优势局限

Redis	内存 KV + 向量一体	极低延迟、缓存/会话/向量统一、运维简单	全内存，超大规模成本高
Milvus	专用分布式向量库	亿级以上、索引类型最全、云原生扩展	组件多、运维较重
Qdrant	Rust 向量库	性能强、过滤好、轻量	生态略小
Weaviate	一体化向量库	模块化、内置混合检索	内存占用偏高
Pinecone	全托管 SaaS	免运维、上手快	闭源、数据上云、成本高
pgvector	PostgreSQL 扩展	复用现有 PG、事务一致	大规模性能一般

8.2 基准参考

Redis 官方基准（2024）：相同召回下，Redis 延迟比 Qdrant 低约 4 倍、比 Milvus 低约 4.67 倍、比 Weaviate 快约 1.71 倍（注意：为厂商自报，口径需自验）。
内存：5 万向量规模，Redis 约 559MB，明显低于多数专用向量库。

8.3 选型决策

#mermaid-svg-2QxqUQKTdONPqDA9{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-2QxqUQKTdONPqDA9 .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-2QxqUQKTdONPqDA9 .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-2QxqUQKTdONPqDA9 .error-icon{fill:#552222;}#mermaid-svg-2QxqUQKTdONPqDA9 .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-2QxqUQKTdONPqDA9 .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-2QxqUQKTdONPqDA9 .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-2QxqUQKTdONPqDA9 .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-2QxqUQKTdONPqDA9 .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-2QxqUQKTdONPqDA9 .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-2QxqUQKTdONPqDA9 .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-2QxqUQKTdONPqDA9 .marker{fill:#333333;stroke:#333333;}#mermaid-svg-2QxqUQKTdONPqDA9 .marker.cross{stroke:#333333;}#mermaid-svg-2QxqUQKTdONPqDA9 svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-2QxqUQKTdONPqDA9 p{margin:0;}#mermaid-svg-2QxqUQKTdONPqDA9 .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-2QxqUQKTdONPqDA9 .cluster-label text{fill:#333;}#mermaid-svg-2QxqUQKTdONPqDA9 .cluster-label span{color:#333;}#mermaid-svg-2QxqUQKTdONPqDA9 .cluster-label span p{background-color:transparent;}#mermaid-svg-2QxqUQKTdONPqDA9 .label text,#mermaid-svg-2QxqUQKTdONPqDA9 span{fill:#333;color:#333;}#mermaid-svg-2QxqUQKTdONPqDA9 .node rect,#mermaid-svg-2QxqUQKTdONPqDA9 .node circle,#mermaid-svg-2QxqUQKTdONPqDA9 .node ellipse,#mermaid-svg-2QxqUQKTdONPqDA9 .node polygon,#mermaid-svg-2QxqUQKTdONPqDA9 .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-2QxqUQKTdONPqDA9 .rough-node .label text,#mermaid-svg-2QxqUQKTdONPqDA9 .node .label text,#mermaid-svg-2QxqUQKTdONPqDA9 .image-shape .label,#mermaid-svg-2QxqUQKTdONPqDA9 .icon-shape .label{text-anchor:middle;}#mermaid-svg-2QxqUQKTdONPqDA9 .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-2QxqUQKTdONPqDA9 .rough-node .label,#mermaid-svg-2QxqUQKTdONPqDA9 .node .label,#mermaid-svg-2QxqUQKTdONPqDA9 .image-shape .label,#mermaid-svg-2QxqUQKTdONPqDA9 .icon-shape .label{text-align:center;}#mermaid-svg-2QxqUQKTdONPqDA9 .node.clickable{cursor:pointer;}#mermaid-svg-2QxqUQKTdONPqDA9 .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-2QxqUQKTdONPqDA9 .arrowheadPath{fill:#333333;}#mermaid-svg-2QxqUQKTdONPqDA9 .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-2QxqUQKTdONPqDA9 .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-2QxqUQKTdONPqDA9 .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-2QxqUQKTdONPqDA9 .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-2QxqUQKTdONPqDA9 .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-2QxqUQKTdONPqDA9 .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-2QxqUQKTdONPqDA9 .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-2QxqUQKTdONPqDA9 .cluster text{fill:#333;}#mermaid-svg-2QxqUQKTdONPqDA9 .cluster span{color:#333;}#mermaid-svg-2QxqUQKTdONPqDA9 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-2QxqUQKTdONPqDA9 .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-2QxqUQKTdONPqDA9 rect.text{fill:none;stroke-width:0;}#mermaid-svg-2QxqUQKTdONPqDA9 .icon-shape,#mermaid-svg-2QxqUQKTdONPqDA9 .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-2QxqUQKTdONPqDA9 .icon-shape p,#mermaid-svg-2QxqUQKTdONPqDA9 .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-2QxqUQKTdONPqDA9 .icon-shape .label rect,#mermaid-svg-2QxqUQKTdONPqDA9 .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-2QxqUQKTdONPqDA9 .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-2QxqUQKTdONPqDA9 .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-2QxqUQKTdONPqDA9 :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

百万级以内，已在用 Redis，要低延迟

亿级以上、专用向量、需弹性扩展

已重度使用 PostgreSQL

完全不想运维、接受上云

是

否，需复杂混合查询

数据规模与需求？

选 Redis

选 Milvus

选 pgvector

选 Pinecone

只需简单相似度？

用 Vector Sets

用 Query Engine

九、RAG 完整架构

#mermaid-svg-PNJGmz4BsCCB4Uiu{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-PNJGmz4BsCCB4Uiu .edge-animation-slow{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 50s linear infinite;stroke-linecap:round;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edge-animation-fast{stroke-dasharray:9,5!important;stroke-dashoffset:900;animation:dash 20s linear infinite;stroke-linecap:round;}#mermaid-svg-PNJGmz4BsCCB4Uiu .error-icon{fill:#552222;}#mermaid-svg-PNJGmz4BsCCB4Uiu .error-text{fill:#552222;stroke:#552222;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edge-thickness-normal{stroke-width:1px;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edge-thickness-thick{stroke-width:3.5px;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edge-pattern-solid{stroke-dasharray:0;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edge-thickness-invisible{stroke-width:0;fill:none;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edge-pattern-dashed{stroke-dasharray:3;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edge-pattern-dotted{stroke-dasharray:2;}#mermaid-svg-PNJGmz4BsCCB4Uiu .marker{fill:#333333;stroke:#333333;}#mermaid-svg-PNJGmz4BsCCB4Uiu .marker.cross{stroke:#333333;}#mermaid-svg-PNJGmz4BsCCB4Uiu svg{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;font-size:16px;}#mermaid-svg-PNJGmz4BsCCB4Uiu p{margin:0;}#mermaid-svg-PNJGmz4BsCCB4Uiu .label{font-family:\”trebuchet ms\”,verdana,arial,sans-serif;color:#333;}#mermaid-svg-PNJGmz4BsCCB4Uiu .cluster-label text{fill:#333;}#mermaid-svg-PNJGmz4BsCCB4Uiu .cluster-label span{color:#333;}#mermaid-svg-PNJGmz4BsCCB4Uiu .cluster-label span p{background-color:transparent;}#mermaid-svg-PNJGmz4BsCCB4Uiu .label text,#mermaid-svg-PNJGmz4BsCCB4Uiu span{fill:#333;color:#333;}#mermaid-svg-PNJGmz4BsCCB4Uiu .node rect,#mermaid-svg-PNJGmz4BsCCB4Uiu .node circle,#mermaid-svg-PNJGmz4BsCCB4Uiu .node ellipse,#mermaid-svg-PNJGmz4BsCCB4Uiu .node polygon,#mermaid-svg-PNJGmz4BsCCB4Uiu .node path{fill:#ECECFF;stroke:#9370DB;stroke-width:1px;}#mermaid-svg-PNJGmz4BsCCB4Uiu .rough-node .label text,#mermaid-svg-PNJGmz4BsCCB4Uiu .node .label text,#mermaid-svg-PNJGmz4BsCCB4Uiu .image-shape .label,#mermaid-svg-PNJGmz4BsCCB4Uiu .icon-shape .label{text-anchor:middle;}#mermaid-svg-PNJGmz4BsCCB4Uiu .node .katex path{fill:#000;stroke:#000;stroke-width:1px;}#mermaid-svg-PNJGmz4BsCCB4Uiu .rough-node .label,#mermaid-svg-PNJGmz4BsCCB4Uiu .node .label,#mermaid-svg-PNJGmz4BsCCB4Uiu .image-shape .label,#mermaid-svg-PNJGmz4BsCCB4Uiu .icon-shape .label{text-align:center;}#mermaid-svg-PNJGmz4BsCCB4Uiu .node.clickable{cursor:pointer;}#mermaid-svg-PNJGmz4BsCCB4Uiu .root .anchor path{fill:#333333!important;stroke-width:0;stroke:#333333;}#mermaid-svg-PNJGmz4BsCCB4Uiu .arrowheadPath{fill:#333333;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edgePath .path{stroke:#333333;stroke-width:2.0px;}#mermaid-svg-PNJGmz4BsCCB4Uiu .flowchart-link{stroke:#333333;fill:none;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edgeLabel{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-PNJGmz4BsCCB4Uiu .edgeLabel p{background-color:rgba(232,232,232, 0.8);}#mermaid-svg-PNJGmz4BsCCB4Uiu .edgeLabel rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-PNJGmz4BsCCB4Uiu .labelBkg{background-color:rgba(232, 232, 232, 0.5);}#mermaid-svg-PNJGmz4BsCCB4Uiu .cluster rect{fill:#ffffde;stroke:#aaaa33;stroke-width:1px;}#mermaid-svg-PNJGmz4BsCCB4Uiu .cluster text{fill:#333;}#mermaid-svg-PNJGmz4BsCCB4Uiu .cluster span{color:#333;}#mermaid-svg-PNJGmz4BsCCB4Uiu 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-PNJGmz4BsCCB4Uiu .flowchartTitleText{text-anchor:middle;font-size:18px;fill:#333;}#mermaid-svg-PNJGmz4BsCCB4Uiu rect.text{fill:none;stroke-width:0;}#mermaid-svg-PNJGmz4BsCCB4Uiu .icon-shape,#mermaid-svg-PNJGmz4BsCCB4Uiu .image-shape{background-color:rgba(232,232,232, 0.8);text-align:center;}#mermaid-svg-PNJGmz4BsCCB4Uiu .icon-shape p,#mermaid-svg-PNJGmz4BsCCB4Uiu .image-shape p{background-color:rgba(232,232,232, 0.8);padding:2px;}#mermaid-svg-PNJGmz4BsCCB4Uiu .icon-shape .label rect,#mermaid-svg-PNJGmz4BsCCB4Uiu .image-shape .label rect{opacity:0.5;background-color:rgba(232,232,232, 0.8);fill:rgba(232,232,232, 0.8);}#mermaid-svg-PNJGmz4BsCCB4Uiu .label-icon{display:inline-block;height:1em;overflow:visible;vertical-align:-0.125em;}#mermaid-svg-PNJGmz4BsCCB4Uiu .node .label-icon path{fill:currentColor;stroke:revert;stroke-width:revert;}#mermaid-svg-PNJGmz4BsCCB4Uiu :root{–mermaid-font-family:\”trebuchet ms\”,verdana,arial,sans-serif;}

在线：问答检索

用户提问

问题向量化

Redis 向量检索 TopK（可加过滤）

拼接相关上下文

大模型生成答案

返回答案

离线：知识库构建

原始文档

切分 Chunk

Embedding 向量化

写入 Redis 向量索引

配套最佳实践：

文档切分要合理（如 500 到 1000 字一块，带重叠）。
检索 TopK 通常 3 到 8，可加"标题/来源/时间"等元数据过滤。
可加一层语义缓存（用 Vector Sets 存历史问答，相似度高于阈值直接返回），降低大模型成本。

十、总结与面试要点

10.1 核心结论

Redis 8 已把向量检索并入开源核心，提供 Query Engine（功能全）和 Vector Sets（轻量省内存）两套能力。

原理链路：内容 → Embedding 向量 → 建索引（FLAT/HNSW）→ 按距离（L2/IP/COSINE）找最近邻。

HNSW 是首选：多层图 + 贪心逐层下降，O(log N)，召回 95% 到 99%。

文本语义默认 COSINE + FLOAT32；省内存用 Q8 量化。

召回调 EF_RUNTIME，内存用量化，规模超亿级考虑 Milvus。

10.2 面试高频问题

向量检索为什么比关键词强？语义相近则向量距离近，不受字面限制。
HNSW 原理？多层图，顶层稀疏快速定位、底层全量精确，贪心逐层下降。
HNSW 参数？ M（连接数）、EF_CONSTRUCTION（建图）、EF_RUNTIME（查询）。
距离度量怎么选？文本 COSINE，距离直观用 L2，归一化后 IP 等价 COSINE。
ANN 会不会不准？会牺牲少量召回（约 1% 到 5%），但快几个数量级。
Redis 向量库的优势？极低延迟、与缓存/会话统一、运维简单；不适合纯内存扛不住的超大规模。