前言大模型时代检索增强生成RAG、智能推荐、多模态检索等场景已成为业务创新的核心方向而向量数据库正是支撑这些场景的底层基石。很多开发者提起向量数据库第一反应是Milvus、Pinecone这类专业组件却忽略了日常高频使用的Redis已经具备了完整的企业级向量数据库能力。无需额外部署新组件无需重构现有技术栈就能基于Redis快速搭建低延迟、高可用的向量检索服务。一、Redis向量数据库核心认知1.1 核心定义Redis向量数据库的能力由Redis官方的RediSearch模块提供核心支撑Redis Stack默认集成了该模块与RedisJSON等扩展组件。它在Redis原生的高性能KV存储能力之上扩展了稠密向量的索引构建、存储、相似度检索与混合查询能力能够将文本、图片、音频、用户画像等非结构化数据转换的向量进行高效的近邻检索。1.2 核心能力边界Redis向量数据库的核心能力覆盖以下维度支持FLOAT32类型的稠密向量最大支持32768维向量提供两种核心索引类型FLAT精确索引与HNSW近似索引支持3种主流相似度度量算法L2欧几里得距离、内积IP、余弦相似度支持向量检索与标签、数值、文本关键词的混合过滤查询完全兼容Redis原生的主从、哨兵、集群高可用架构兼容Redis原生的持久化机制RDB/AOF、事务、过期策略等所有能力1.3 选型对比Redis向量库 vs 专业向量数据库对比维度Redis向量数据库专业向量数据库Milvus/Pinecone部署成本极低基于现有Redis Stack即可部署无额外组件较高需独立部署分布式集群依赖对象存储、协调服务等组件学习成本极低兼容Redis原生协议与Java客户端复用现有技术栈较高需学习独立的API、索引配置与集群运维体系数据规模适配百万级到千万级向量数据适配亿级以上超大规模向量数据支持分布式分片与水平扩展检索延迟亚毫秒级Redis原生内存级访问优势毫秒级分布式架构带来一定的网络开销高级特性满足基础向量检索与混合查询需求支持向量压缩、多租户、动态schema、多模态向量、增量索引等高级特性业务适配中小规模数据量、快速迭代的业务已有Redis技术栈的团队超大规模数据、复杂检索场景、企业级多租户需求的业务二、底层原理深度拆解2.1 向量的本质与高维空间语义向量的本质是将非结构化数据转换为一组固定长度的浮点数数组数组中的每一个数值都对应着数据的一个底层特征。比如一段文本“Java并发编程原理”通过Embedding模型转换为1024维的float数组每一个数值都代表了文本的语义特征。在高维空间中语义、特征越相似的数据对应的向量点之间的距离就越近。向量检索的核心逻辑就是计算查询向量与库中所有向量的相似度返回距离最近的TopK个结果也就是常说的K近邻KNN检索。2.2 Redis向量的存储结构Redis向量支持两种存储载体均能被RediSearch索引识别与检索Hash结构向量字段以BLOB二进制格式存储在Hash的field中是性能最优、内存占用最低的存储方式。FLOAT32类型的向量每个维度占用4个字节1024维的向量仅占用4KB内存存储效率极高。JSON结构向量以浮点数组的形式存储在RedisJSON文档中适合需要同时存储向量与复杂结构化数据的场景内存占用略高于Hash结构。无论哪种存储方式向量数据都完全兼容Redis原生的持久化、主从同步、集群分片机制无需额外的存储适配。2.3 核心索引类型与底层实现索引是向量检索性能的核心Redis提供两种索引类型分别适配不同的业务场景。2.3.1 FLAT索引精确KNN检索FLAT索引是暴力搜索索引底层采用全量扫描的方式计算查询向量与库中所有向量的相似度最终返回距离最近的TopK结果。核心优势召回率100%无任何精度损失支持向量的实时增删改数据更新后立即生效核心劣势数据量超过10万条后检索延迟会显著上升不适合大数据量场景适用场景万级以下小数据集、需要100%精确匹配的场景、数据更新频率极高的场景2.3.2 HNSW索引近似ANN检索HNSW全称Hierarchical Navigable Small World即层次化导航小世界图是目前工业界最主流的近似近邻检索算法。底层原理将高维向量构建成多层有向图结构最底层包含所有向量节点越往上的层级节点数量越少。检索时从最顶层的入口节点开始逐层向下查找距离最近的节点最终在最底层找到TopK结果避免了全量扫描。核心优势检索延迟极低百万级向量数据可实现亚毫秒级响应适合大数据量高并发场景核心劣势有轻微的召回率损失索引构建有一定的资源开销数据实时更新性能弱于FLAT索引适用场景十万级以上大数据集、高并发检索场景、可接受轻微精度损失换取极致性能的业务2.4 相似度度量算法全解析Redis支持3种主流的相似度度量算法分别适配不同的业务场景算法的选择直接影响检索效果。2.4.1 L2欧几里得距离L2距离计算的是高维空间中两个向量点之间的直线距离计算公式为距离取值范围[0, ∞)值越小向量相似度越高适用场景图像特征匹配、语音识别、用户行为特征匹配等对向量绝对值敏感的场景2.4.2 内积IP内积计算的是两个向量的点积计算公式为取值范围(-∞, ∞)值越大向量相似度越高核心特性对向量的长度敏感向量模长越大内积值越高适用场景已做L2归一化的向量检索、推荐系统中的用户-物品匹配场景2.4.3 余弦相似度余弦相似度计算的是两个向量夹角的余弦值衡量的是向量的方向相似度与向量长度无关计算公式为取值范围[-1, 1]值越大向量相似度越高1代表完全相同核心特性不受向量模长影响仅关注向量的方向完美适配文本语义匹配适用场景文本语义检索、RAG场景、文档相似度匹配等NLP相关场景❝关键知识点若向量已完成L2归一化模长为1则余弦相似度与内积的计算结果完全一致可直接使用内积算法减少计算开销。三、环境搭建与项目整合3.1 环境快速搭建Redis向量能力依赖RediSearch模块推荐使用Redis Stack官方镜像一键部署Docker部署命令如下docker run -d --name redis-stack -p 6379:6379 -p 8001:8001 redis/redis-stack:7.4.0-v06379端口Redis原生服务端口8001端口Redis Insight可视化管理界面可直接查看向量索引与数据3.2 核心基础命令先通过Redis CLI掌握核心命令理解向量操作的核心逻辑创建FLAT向量索引FT.CREATE idx_doc_flat ON HASH PREFIX 1 doc: SCHEMA content TEXT tags TAG vector VECTOR FLAT 6 TYPE FLOAT32 DIM 1024 DISTANCE_METRIC COSINE2.创建HNSW向量索引FT.CREATE idx_doc_hnsw ON HASH PREFIX 1 doc: SCHEMA content TEXT tags TAG vector VECTOR HNSW 10 TYPE FLOAT32 DIM 1024 DISTANCE_METRIC COSINE M 16 EF_CONSTRUCTION 200 EF_RUNTIME 103.写入向量数据HSET doc:001 content Java并发编程核心原理 tags Java,后端,并发 vector 向量二进制字节数组4.KNN向量检索FT.SEARCH idx_doc_hnsw *[KNN 5 vector $query_vec] PARAMS 2 query_vec 查询向量二进制字节数组 SORTBY __vector_score LIMIT 0 53.3 Java项目环境整合基于Spring Boot 3.2.x搭建项目pom.xml配置如下?xml version1.0 encodingUTF-8? project xmlnshttp://maven.apache.org/POM/4.0.0 xmlns:xsihttp://www.w3.org/2001/XMLSchema-instance xsi:schemaLocationhttp://maven.apache.org/POM/4.0.0 https://maven.apache.org/xsd/maven-4.0.0.xsd modelVersion4.0.0/modelVersion parent groupIdorg.springframework.boot/groupId artifactIdspring-boot-starter-parent/artifactId version3.2.5/version relativePath/ /parent groupIdcom.jam.demo/groupId artifactIdredis-vector-demo/artifactId version0.0.1-SNAPSHOT/version nameredis-vector-demo/name properties java.version17/java.version jedis.version5.1.5/jedis.version fastjson2.version2.0.52/fastjson2.version guava.version33.1.0-jre/guava.version mybatis-plus.version3.5.7/mybatis-plus.version springdoc.version2.5.0/springdoc.version /properties dependencies dependency groupIdorg.springframework.boot/groupId artifactIdspring-boot-starter-web/artifactId /dependency dependency groupIdorg.springframework.boot/groupId artifactIdspring-boot-starter-data-redis/artifactId /dependency dependency groupIdredis.clients/groupId artifactIdjedis/artifactId version${jedis.version}/version /dependency dependency groupIdorg.springdoc/groupId artifactIdspringdoc-openapi-starter-webmvc-ui/artifactId version${springdoc.version}/version /dependency dependency groupIdcom.baomidou/groupId artifactIdmybatis-plus-spring-boot3-starter/artifactId version${mybatis-plus.version}/version /dependency dependency groupIdcom.mysql/groupId artifactIdmysql-connector-j/artifactId scoperuntime/scope /dependency dependency groupIdcom.alibaba.fastjson2/groupId artifactIdfastjson2/artifactId version${fastjson2.version}/version /dependency dependency groupIdcom.google.guava/groupId artifactIdguava/artifactId version${guava.version}/version /dependency dependency groupIdorg.projectlombok/groupId artifactIdlombok/artifactId version1.18.32/version scopeprovided/scope /dependency dependency groupIdorg.springframework.boot/groupId artifactIdspring-boot-starter-test/artifactId scopetest/scope /dependency /dependencies build plugins plugin groupIdorg.springframework.boot/groupId artifactIdspring-boot-maven-plugin/artifactId configuration excludes exclude groupIdorg.projectlombok/groupId artifactIdlombok/artifactId /exclude /excludes /configuration /plugin /plugins /build /projectapplication.yml配置文件spring: application: name: redis-vector-demo data: redis: host: 127.0.0.1 port: 6379 database: 0 timeout: 5000 jedis: pool: max-active: 20 max-idle: 10 min-idle: 5 max-wait: 3000 datasource: driver-class-name: com.mysql.cj.jdbc.Driver url: jdbc:mysql://127.0.0.1:3306/vector_db?useUnicodetruecharacterEncodingutf8serverTimezoneAsia/ShanghaiuseSSLfalse username: root password: root mybatis-plus: configuration: map-underscore-to-camel-case: true log-impl: org.apache.ibatis.logging.stdout.StdOutImpl global-config: db-config: id-type: auto logic-delete-field: deleted logic-delete-value: 1 logic-not-delete-value: 0 springdoc: swagger-ui: path: /swagger-ui.html enabled: true api-docs: enabled: true path: /v3/api-docsRedis配置类package com.jam.demo.config; import org.springframework.context.annotation.Bean; import org.springframework.context.annotation.Configuration; import org.springframework.data.redis.connection.RedisConnectionFactory; import org.springframework.data.redis.connection.jedis.JedisConnectionFactory; import org.springframework.data.redis.core.RedisTemplate; import org.springframework.data.redis.serializer.StringRedisSerializer; import redis.clients.jedis.JedisPool; import redis.clients.jedis.JedisPoolConfig; /** * Redis配置类 * author ken * since 2026-04-03 */ Configuration public class RedisConfig { Bean public RedisTemplateString, Object redisTemplate(RedisConnectionFactory connectionFactory) { RedisTemplateString, Object redisTemplate new RedisTemplate(); redisTemplate.setConnectionFactory(connectionFactory); StringRedisSerializer stringRedisSerializer new StringRedisSerializer(); redisTemplate.setKeySerializer(stringRedisSerializer); redisTemplate.setHashKeySerializer(stringRedisSerializer); redisTemplate.setValueSerializer(stringRedisSerializer); redisTemplate.setHashValueSerializer(stringRedisSerializer); redisTemplate.afterPropertiesSet(); return redisTemplate; } Bean public JedisPool jedisPool(JedisConnectionFactory jedisConnectionFactory) { JedisPoolConfig poolConfig new JedisPoolConfig(); poolConfig.setMaxTotal(20); poolConfig.setMaxIdle(10); poolConfig.setMinIdle(5); poolConfig.setMaxWaitMillis(3000); return new JedisPool(poolConfig, jedisConnectionFactory.getHostName(), jedisConnectionFactory.getPort(), jedisConnectionFactory.getTimeout(), jedisConnectionFactory.getPassword(), jedisConnectionFactory.getDatabase()); } }向量工具类核心解决Java大端序与Redis要求的小端序转换问题package com.jam.demo.util; import org.springframework.util.ObjectUtils; import java.nio.ByteBuffer; import java.nio.ByteOrder; /** * 向量处理工具类 * author ken * since 2026-04-03 */ public class VectorUtils { private VectorUtils() {} /** * float向量数组转换为Redis要求的小端序字节数组 * param vector 浮点向量数组 * return 小端序字节数组 */ public static byte[] floatVectorToLittleEndianBytes(float[] vector) { if (ObjectUtils.isEmpty(vector)) { return new byte[0]; } ByteBuffer buffer ByteBuffer.allocate(Float.BYTES * vector.length); buffer.order(ByteOrder.LITTLE_ENDIAN); for (float value : vector) { buffer.putFloat(value); } return buffer.array(); } /** * Redis小端序字节数组转换为float向量数组 * param bytes 小端序字节数组 * return 浮点向量数组 */ public static float[] littleEndianBytesToFloatVector(byte[] bytes) { if (ObjectUtils.isEmpty(bytes)) { return new float[0]; } int floatCount bytes.length / Float.BYTES; float[] vector new float[floatCount]; ByteBuffer buffer ByteBuffer.wrap(bytes); buffer.order(ByteOrder.LITTLE_ENDIAN); for (int i 0; i floatCount; i) { vector[i] buffer.getFloat(); } return vector; } /** * 向量L2归一化处理 * param vector 原始向量 * return 归一化后的向量 */ public static float[] normalizeVector(float[] vector) { if (ObjectUtils.isEmpty(vector)) { return new float[0]; } double sum 0.0; for (float value : vector) { sum Math.pow(value, 2); } float norm (float) Math.sqrt(sum); if (norm 0) { return vector; } float[] normalizedVector new float[vector.length]; for (int i 0; i vector.length; i) { normalizedVector[i] vector[i] / norm; } return normalizedVector; } }四、核心功能全实战4.1 向量索引的创建与管理实体类定义package com.jam.demo.entity; import com.baomidou.mybatisplus.annotation.IdType; import com.baomidou.mybatisplus.annotation.TableId; import com.baomidou.mybatisplus.annotation.TableName; import io.swagger.v3.oas.annotations.media.Schema; import lombok.Data; import java.time.LocalDateTime; /** * 向量文档实体 * author ken * since 2026-04-03 */ Data TableName(t_vector_document) Schema(description 向量文档实体) public class VectorDocument { TableId(type IdType.AUTO) Schema(description 主键ID, example 1) private Long id; Schema(description 文档唯一ID, example doc:001) private String docId; Schema(description 文档内容, example Java并发编程核心原理) private String content; Schema(description 标签逗号分隔, example Java,后端,并发) private String tags; Schema(description 向量维度, example 1024) private Integer vectorDimension; Schema(description 创建时间) private LocalDateTime createTime; Schema(description 更新时间) private LocalDateTime updateTime; }Mapper接口package com.jam.demo.mapper; import com.baomidou.mybatisplus.core.mapper.BaseMapper; import com.jam.demo.entity.VectorDocument; import org.apache.ibatis.annotations.Mapper; /** * 向量文档Mapper * author ken * since 2026-04-03 */ Mapper public interface VectorDocumentMapper extends BaseMapperVectorDocument { }索引管理Service接口package com.jam.demo.service; import com.jam.demo.entity.VectorDocument; import com.jam.demo.vo.VectorSearchResult; import java.util.List; /** * Redis向量服务接口 * author ken * since 2026-04-03 */ public interface RedisVectorService { boolean createFlatIndex(String indexName, int dimension, String distanceMetric); boolean createHnswIndex(String indexName, int dimension, String distanceMetric, int m, int efConstruction, int efRuntime); boolean deleteIndex(String indexName); boolean existIndex(String indexName); boolean saveVectorDocument(VectorDocument document, float[] vector); boolean deleteVectorDocument(String docId); ListVectorSearchResult knnSearch(String indexName, float[] queryVector, int topK); ListVectorSearchResult hybridSearch(String indexName, float[] queryVector, int topK, String tagFilter); }索引管理Service实现类核心方法package com.jam.demo.service.impl; import com.jam.demo.entity.VectorDocument; import com.jam.demo.mapper.VectorDocumentMapper; import com.jam.demo.service.RedisVectorService; import com.jam.demo.util.VectorUtils; import com.jam.demo.vo.VectorSearchResult; import lombok.extern.slf4j.Slf4j; import org.springframework.stereotype.Service; import org.springframework.transaction.TransactionStatus; import org.springframework.transaction.support.TransactionCallback; import org.springframework.transaction.support.TransactionTemplate; import org.springframework.util.CollectionUtils; import org.springframework.util.ObjectUtils; import org.springframework.util.StringUtils; import redis.clients.jedis.Jedis; import redis.clients.jedis.JedisPool; import redis.clients.jedis.search.Document; import redis.clients.jedis.search.Query; import redis.clients.jedis.search.SearchResult; import jakarta.annotation.Resource; import java.util.ArrayList; import java.util.List; import java.util.Map; /** * Redis向量服务实现类 * author ken * since 2026-04-03 */ Slf4j Service public class RedisVectorServiceImpl implements RedisVectorService { private static final String DOC_KEY_PREFIX doc:; private static final String VECTOR_FIELD_NAME vector; private static final String CONTENT_FIELD_NAME content; private static final String TAGS_FIELD_NAME tags; private static final String SCORE_FIELD_NAME __vector_score; Resource private JedisPool jedisPool; Resource private VectorDocumentMapper vectorDocumentMapper; Resource private TransactionTemplate transactionTemplate; Override public boolean createFlatIndex(String indexName, int dimension, String distanceMetric) { if (!StringUtils.hasText(indexName) || dimension 0 || !StringUtils.hasText(distanceMetric)) { log.error(创建FLAT索引参数异常); return false; } try (Jedis jedis jedisPool.getResource()) { String schema String.format( ON HASH PREFIX 1 %s SCHEMA %s TEXT %s TAG %s VECTOR FLAT 6 TYPE FLOAT32 DIM %d DISTANCE_METRIC %s, DOC_KEY_PREFIX, CONTENT_FIELD_NAME, TAGS_FIELD_NAME, VECTOR_FIELD_NAME, dimension, distanceMetric ); jedis.ftCreate(indexName, schema.split( )); log.info(FLAT索引创建成功索引名{}, indexName); return true; } catch (Exception e) { log.error(FLAT索引创建失败索引名{}, indexName, e); return false; } } Override public boolean createHnswIndex(String indexName, int dimension, String distanceMetric, int m, int efConstruction, int efRuntime) { if (!StringUtils.hasText(indexName) || dimension 0 || !StringUtils.hasText(distanceMetric)) { log.error(创建HNSW索引参数异常); return false; } try (Jedis jedis jedisPool.getResource()) { String schema String.format( ON HASH PREFIX 1 %s SCHEMA %s TEXT %s TAG %s VECTOR HNSW 10 TYPE FLOAT32 DIM %d DISTANCE_METRIC %s M %d EF_CONSTRUCTION %d EF_RUNTIME %d, DOC_KEY_PREFIX, CONTENT_FIELD_NAME, TAGS_FIELD_NAME, VECTOR_FIELD_NAME, dimension, distanceMetric, m, efConstruction, efRuntime ); jedis.ftCreate(indexName, schema.split( )); log.info(HNSW索引创建成功索引名{}, indexName); return true; } catch (Exception e) { log.error(HNSW索引创建失败索引名{}, indexName, e); return false; } } Override public boolean deleteIndex(String indexName) { if (!StringUtils.hasText(indexName)) { return false; } try (Jedis jedis jedisPool.getResource()) { jedis.ftDropIndex(indexName); log.info(索引删除成功索引名{}, indexName); return true; } catch (Exception e) { log.error(索引删除失败索引名{}, indexName, e); return false; } } Override public boolean existIndex(String indexName) { if (!StringUtils.hasText(indexName)) { return false; } try (Jedis jedis jedisPool.getResource()) { jedis.ftInfo(indexName); return true; } catch (Exception e) { return false; } } }4.2 向量数据的增删改查Service实现类新增增删改查核心方法Override public boolean saveVectorDocument(VectorDocument document, float[] vector) { if (ObjectUtils.isEmpty(document) || !StringUtils.hasText(document.getDocId()) || ObjectUtils.isEmpty(vector)) { log.error(保存向量文档参数异常); return false; } return transactionTemplate.execute(new TransactionCallbackBoolean() { Override public Boolean doInTransaction(TransactionStatus status) { try { int dbResult; VectorDocument existDoc vectorDocumentMapper.selectOne( new com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapperVectorDocument() .eq(VectorDocument::getDocId, document.getDocId()) ); if (ObjectUtils.isEmpty(existDoc)) { dbResult vectorDocumentMapper.insert(document); } else { document.setId(existDoc.getId()); dbResult vectorDocumentMapper.updateById(document); } if (dbResult 0) { status.setRollbackOnly(); log.error(向量文档数据库保存失败docId{}, document.getDocId()); return false; } try (Jedis jedis jedisPool.getResource()) { byte[] vectorBytes VectorUtils.floatVectorToLittleEndianBytes(vector); MapString, String hashData Map.of( CONTENT_FIELD_NAME, document.getContent(), TAGS_FIELD_NAME, ObjectUtils.isEmpty(document.getTags()) ? : document.getTags() ); jedis.hset(DOC_KEY_PREFIX document.getDocId(), hashData); jedis.hset(DOC_KEY_PREFIX document.getDocId(), VECTOR_FIELD_NAME, vectorBytes); } log.info(向量文档保存成功docId{}, document.getDocId()); return true; } catch (Exception e) { status.setRollbackOnly(); log.error(向量文档保存失败docId{}, document.getDocId(), e); return false; } } }); } Override public boolean deleteVectorDocument(String docId) { if (!StringUtils.hasText(docId)) { return false; } return transactionTemplate.execute(new TransactionCallbackBoolean() { Override public Boolean doInTransaction(TransactionStatus status) { try { vectorDocumentMapper.delete( new com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapperVectorDocument() .eq(VectorDocument::getDocId, docId) ); try (Jedis jedis jedisPool.getResource()) { jedis.del(DOC_KEY_PREFIX docId); } log.info(向量文档删除成功docId{}, docId); return true; } catch (Exception e) { status.setRollbackOnly(); log.error(向量文档删除失败docId{}, docId, e); return false; } } }); }检索结果VO定义package com.jam.demo.vo; import io.swagger.v3.oas.annotations.media.Schema; import lombok.Data; /** * 向量检索结果VO * author ken * since 2026-04-03 */ Data Schema(description 向量检索结果VO) public class VectorSearchResult { Schema(description 文档ID, example doc:001) private String docId; Schema(description 文档内容, example Java并发编程核心原理) private String content; Schema(description 文档标签, example Java,后端,并发) private String tags; Schema(description 相似度得分, example 0.92) private Double score; }4.3 精准KNN与近似ANN检索Service实现类新增检索核心方法Override public ListVectorSearchResult knnSearch(String indexName, float[] queryVector, int topK) { if (!StringUtils.hasText(indexName) || ObjectUtils.isEmpty(queryVector) || topK 0) { log.error(KNN检索参数异常); return new ArrayList(); } try (Jedis jedis jedisPool.getResource()) { byte[] queryVectorBytes VectorUtils.floatVectorToLittleEndianBytes(queryVector); String queryString String.format(*[KNN %d %s $query_vec], topK, VECTOR_FIELD_NAME); Query query new Query(queryString) .addParam(query_vec, queryVectorBytes) .setSortBy(SCORE_FIELD_NAME, true) .limit(0, topK) .returnFields(CONTENT_FIELD_NAME, TAGS_FIELD_NAME, SCORE_FIELD_NAME); SearchResult searchResult jedis.ftSearch(indexName, query); ListDocument documents searchResult.getDocuments(); if (CollectionUtils.isEmpty(documents)) { return new ArrayList(); } ListVectorSearchResult resultList new ArrayList(documents.size()); for (Document doc : documents) { VectorSearchResult result new VectorSearchResult(); result.setDocId(doc.getId().replace(DOC_KEY_PREFIX, )); result.setContent(doc.getString(CONTENT_FIELD_NAME)); result.setTags(doc.getString(TAGS_FIELD_NAME)); result.setScore(1 - Double.parseDouble(doc.getString(SCORE_FIELD_NAME))); resultList.add(result); } return resultList; } catch (Exception e) { log.error(KNN检索失败索引名{}, indexName, e); return new ArrayList(); } } Override public ListVectorSearchResult hybridSearch(String indexName, float[] queryVector, int topK, String tagFilter) { if (!StringUtils.hasText(indexName) || ObjectUtils.isEmpty(queryVector) || topK 0 || !StringUtils.hasText(tagFilter)) { log.error(混合检索参数异常); return new ArrayList(); } try (Jedis jedis jedisPool.getResource()) { byte[] queryVectorBytes VectorUtils.floatVectorToLittleEndianBytes(queryVector); String queryString String.format(%s:{%s} [KNN %d %s $query_vec], TAGS_FIELD_NAME, tagFilter, topK, VECTOR_FIELD_NAME); Query query new Query(queryString) .addParam(query_vec, queryVectorBytes) .setSortBy(SCORE_FIELD_NAME, true) .limit(0, topK) .returnFields(CONTENT_FIELD_NAME, TAGS_FIELD_NAME, SCORE_FIELD_NAME); SearchResult searchResult jedis.ftSearch(indexName, query); ListDocument documents searchResult.getDocuments(); if (CollectionUtils.isEmpty(documents)) { return new ArrayList(); } ListVectorSearchResult resultList new ArrayList(documents.size()); for (Document doc : documents) { VectorSearchResult result new VectorSearchResult(); result.setDocId(doc.getId().replace(DOC_KEY_PREFIX, )); result.setContent(doc.getString(CONTENT_FIELD_NAME)); result.setTags(doc.getString(TAGS_FIELD_NAME)); result.setScore(1 - Double.parseDouble(doc.getString(SCORE_FIELD_NAME))); resultList.add(result); } return resultList; } catch (Exception e) { log.error(混合检索失败索引名{}, indexName, e); return new ArrayList(); } }