Merge pull request #315 from MuSan-Li/update/2026-07-19-fixes

修复RAG相关逻辑优化:统一文档切割配置、向量数据生命周期、统一检索服务(阈值/重排/缓存)、按知识库路由向量库、重新解析支持
2026-09-13 00:14:59 +00:00 · 2026-07-21 14:13:42 +08:00
parent 24bee53f9e cf5423f69f
commit 46a8d6b552
68 changed files with 1311 additions and 345 deletions
--- a/.gitignore
+++ b/.gitignore
@@ -53,3 +53,4 @@ logs/

 .flattened-pom.xml
 /.claude/settings.local.json
+/docs/docker/milvus/volumes/
--- a/RAG_TEST_REPORT_2026-07-20.md
+++ b/RAG_TEST_REPORT_2026-07-20.md
@@ -0,0 +1,64 @@
+# RAG 完整修复与全量验收报告
+
+验收时间：2026-07-21（Asia/Shanghai）  
+验收对象：当前未提交工作区（保留原有改动）
+
+## 结论
+
+计划内的 1–15 项工程缺陷已完成代码修复，默认 Maven 构建已从“跳过测试”改为真实执行测试。全仓 37 个 reactor 模块测试成功，`ruoyi-chat` 49/49 通过，两个前端生产构建通过，`git diff --check` 通过。
+
+本机已运行 MySQL、Redis、MinIO 和 Weaviate 1.30.0；Milvus/Qdrant 容器以及有效的 embedding/chat/rerank provider 凭证不存在，因此这三项真实 provider/存储引擎冒烟被标记为环境限制，不影响确定性代码验收。
+
+## 1–15 项验收
+
+| # | 状态 | 修复/证据 |
+|---|---|---|
+| 1 | 通过 | Markdown/Java/字符分片的边界、空文档、超长块回归通过。 |
+| 2 | 通过 | Supervisor 每轮仅保留一个 RAG 入口，不再用已含 RAG 的 prompt 重复检索。 |
+| 3 | 通过 | 历史消息进入 Supervisor prompt，检索 query 与最终 prompt 分离。 |
+| 4 | 通过 | `fid` 稳定 ID 贯穿 DB/三种向量库/RRF，融合去重回归通过。 |
+| 5 | 通过 | aiflow vector/hybrid 复用统一检索服务；graph 明确返回不支持，不再伪装为 vector。 |
+| 6 | 通过 | 重解析改为先写新 fid、再清旧向量、最后替换 DB；失败补偿新向量；删片段/附件/库遇向量删除失败即中止。 |
+| 7 | 通过 | embedding/rerank provider 使用 prototype 实例，工厂缓存可按模型刷新，避免跨配置污染。 |
+| 8 | 通过 | `similarityThreshold` 仅用于粗召回；`rerankScoreThreshold` 仅在 rerank 真实成功后生效，回归测试通过。 |
+| 9 | 通过 | 默认配置和 Compose 统一为 Weaviate 1.30.0、`28080:8080`。 |
+| 10 | 通过 | 三种策略均使用 `embedAll`；Weaviate batch objects、Milvus `addAll`、Qdrant `addAll`。 |
+| 11 | 通过 | upload/parse/retrieval 权限保留，parse/retrieval 增加分布式防重复提交，upload 由现有知识库+文件名唯一约束兜底。 |
+| 12 | 通过 | 分隔符使用字面量语义，`|`/`.`/`*` 回归通过。 |
+| 13 | 通过 | hybrid 通道失败可降级到 vector；所有可用通道都失败时抛出明确业务异常。 |
+| 14 | 通过 | Weaviate client 稳定懒加载单例；schema 仅在已存在或创建成功后进入缓存。 |
+| 15 | 通过 | 工厂新增严格 `getStrategy(type)`，知识库 `vectorModel` 优先，空值才回退全局，非法值直接报错。 |
+
+## 其他完成项
+
+- 多知识库并行检索，按 `kid + docId + fid` 去重，统一上限和字符预算。
+- 5 分钟短 TTL 检索缓存，key 覆盖检索参数，知识数据变更主动失效。
+- rerank 仅保留 provider 实际返回的文档。
+- 知识库文档数改为 group-by 查询，消除该 N+1。
+- Milvus/Qdrant/Weaviate 的删 collection/doc/fid 语义对齐；Milvus 删库改为 drop collection。
+- MCP `npx` 根据操作系统解析，支持系统属性/环境变量覆盖。
+- `fid` 非空唯一、`doc_id varchar(32)`、租户/用户索引与可重复执行迁移脚本已提供。
+- 用户端聊天页已接入知识库列表和最小选择器。
+
+## 测试记录
+
+| 检查 | 结果 |
+|---|---|
+| `mvn -Pdev test` | 37/37 reactor 模块 SUCCESS；`ruoyi-chat` 49/49 |
+| `mvn -Pdev -pl ruoyi-modules/ruoyi-aiflow -am -DskipTests compile` | 21/21 SUCCESS |
+| `ruoyi-web: pnpm build` | SUCCESS，2621 modules transformed |
+| `ruoyi-admin: pnpm build` | SUCCESS，10/10 build tasks |
+| `git diff --check` | SUCCESS，无空白错误 |
+| Weaviate/MySQL/Redis/MinIO | Docker 服务运行，Weaviate 1.30.0 映射 28080 |
+| 三向量库 Docker 集成 | SUCCESS；Weaviate 1.30.0、Milvus 2.5.7、Qdrant 1.17.0 真实写入/检索/删除测试 3/3 通过 |
+| 真实 embedding/chat/rerank | 环境限制：当前配置为无效/占位凭证 |
+
+本轮未创建新的 `codex_rag_verify_` 持久化数据；上一轮验收数据已清理，未动现有非测试数据。
+
+## 2026-07-21 三向量库 Docker 补充验收
+
+- 启动并保留 `ruoyi-rag-milvus`、`ruoyi-rag-milvus-etcd`、`ruoyi-rag-milvus-minio`、`ruoyi-rag-qdrant`，四个容器健康检查均为 `healthy`。
+- Milvus 专用 MinIO 仅在 Docker 内网可达，没有占用宿主机 9000/9001；Milvus 映射 19530/9091，Qdrant 映射 6333/6334。
+- `ThreeVectorStoresDockerIT` 使用 32 维确定性 embedding，对三库逐一验证 batch write、vector search、fid delete、docId delete 和 drop collection，3/3 通过。
+- 首轮测试发现 Milvus `autoFlush=false` 导致批量写入后不可立即检索、元数据删除不可立即见；改为写入和删除返回前 flush 后通过。
+- 清理后 Weaviate/Qdrant 的 `CodexRagVerify*` collection 计数均为 0，Milvus collection 也由测试 finally 成功 drop；本轮未写入 MySQL 或 OSS 测试数据。
--- a/docs/docker/milvus/docker-compose.yml
+++ b/docs/docker/milvus/docker-compose.yml
@@ -1,8 +1,6 @@
-version: '3.5'
- 
 services:
  etcd:
-    container_name: milvus-etcd
+    container_name: ruoyi-rag-milvus-etcd
    image: quay.io/coreos/etcd:v3.5.18
    environment:
      - ETCD_AUTO_COMPACTION_MODE=revision
@@ -19,14 +17,11 @@ services:
      retries: 3
 
  minio:
-    container_name: milvus-minio
+    container_name: ruoyi-rag-milvus-minio
    image: minio/minio:RELEASE.2023-03-20T20-16-18Z
    environment:
      MINIO_ACCESS_KEY: minioadmin
      MINIO_SECRET_KEY: minioadmin
-    ports:
-      - "9001:9001"
-      - "9000:9000"
    volumes:
      - ${DOCKER_VOLUME_DIRECTORY:-.}/volumes/minio:/minio_data
    command: minio server /minio_data --console-address ":9001"
@@ -37,7 +32,7 @@ services:
      retries: 3
 
  standalone:
-    container_name: milvus-standalone
+    container_name: ruoyi-rag-milvus
    image: milvusdb/milvus:v2.5.7
    command: ["milvus", "run", "standalone"]
    security_opt:
@@ -61,7 +56,7 @@ services:
      - "minio"
 
  attu:
-    container_name: attu
+    container_name: ruoyi-rag-attu
    image: zilliz/attu:v2.5.7
    environment:
      MILVUS_URL: milvus-standalone:19530
@@ -72,4 +67,4 @@ services:
 
 networks:
  default:
-    name: milvus
+    name: ruoyi-rag-milvus
--- a/docs/docker/qdrant/docker-compose.yml
+++ b/docs/docker/qdrant/docker-compose.yml
@@ -1,12 +1,20 @@
 ---
 services:
  qdrant:
-    image: qdrant/qdrant:latest
+    container_name: ruoyi-rag-qdrant
+    image: qdrant/qdrant:v1.17.0
    ports:
      - 6333:6333
      - 6334:6334
    volumes:
      - qdrant_data:/qdrant/storage
+    healthcheck:
+      test: ["CMD-SHELL", "bash -c 'exec 3<>/dev/tcp/127.0.0.1/6333 && printf \"GET /healthz HTTP/1.1\\r\\nHost: localhost\\r\\nConnection: close\\r\\n\\r\\n\" >&3 && grep -q \"200 OK\" <&3'"]
+      interval: 10s
+      timeout: 5s
+      retries: 12
+      start_period: 10s
+    restart: unless-stopped
 volumes:
  qdrant_data:
 ...
--- a/docs/docker/weaviate/docker-compose.yml
+++ b/docs/docker/weaviate/docker-compose.yml
@@ -5,12 +5,12 @@ services:
      - --host
      - 0.0.0.0
      - --port
-      - '6038'
+      - '8080'
      - --scheme
      - http
-    image: semitechnologies/weaviate:1.19.7
+    image: semitechnologies/weaviate:1.30.0
    ports:
-      - 6038:6038
+      - 28080:8080
      - 50051:50051
    volumes:
      - weaviate_data:/var/lib/weaviate
--- a/docs/script/sql/ruoyi-ai.sql
+++ b/docs/script/sql/ruoyi-ai.sql
@@ -1131,7 +1131,8 @@ CREATE TABLE `knowledge_attach`  (
                                     `id` bigint NOT NULL AUTO_INCREMENT COMMENT '主键',
                                     `knowledge_id` bigint NOT NULL COMMENT '知识库ID',
                                     `oss_id` bigint NULL DEFAULT NULL COMMENT '对象存储ID',
-                                     `doc_id` varchar(11) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL DEFAULT NULL COMMENT '文档ID',
+                                     `doc_id` varchar(32) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL DEFAULT NULL COMMENT '文档ID',
+                                     `file_hash` varchar(64) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL DEFAULT NULL COMMENT '文件SHA-256摘要',
                                     `name` varchar(500) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL DEFAULT NULL COMMENT '附件名称',
                                     `type` varchar(50) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NOT NULL COMMENT '附件类型',
                                     `create_dept` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL DEFAULT NULL COMMENT '部门',
@@ -1143,7 +1144,8 @@ CREATE TABLE `knowledge_attach`  (
                                     `tenant_id` bigint NOT NULL DEFAULT 0 COMMENT '租户Id',
                                     `status` tinyint NULL DEFAULT 0 COMMENT '解析状态: 0待解析, 1解析中, 2已解析, 3解析失败',
                                     PRIMARY KEY (`id`) USING BTREE,
-                                     UNIQUE INDEX `idx_kname`(`knowledge_id` ASC, `name` ASC) USING BTREE
+                                     UNIQUE INDEX `idx_kname`(`knowledge_id` ASC, `name` ASC) USING BTREE,
+                                     UNIQUE INDEX `uk_knowledge_file_hash`(`knowledge_id`, `file_hash`) USING BTREE
 ) ENGINE = InnoDB AUTO_INCREMENT = 2033199209203183619 CHARACTER SET = utf8mb4 COLLATE = utf8mb4_0900_ai_ci COMMENT = '知识库附件' ROW_FORMAT = DYNAMIC;

 -- ----------------------------
@@ -1156,8 +1158,9 @@ CREATE TABLE `knowledge_attach`  (
 DROP TABLE IF EXISTS `knowledge_fragment`;
 CREATE TABLE `knowledge_fragment`  (
                                       `id` bigint NOT NULL AUTO_INCREMENT COMMENT '主键',
+                                       `fid` varchar(32) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NOT NULL COMMENT '向量库片段ID',
                                       `idx` int NOT NULL COMMENT '片段索引下标',
-                                       `doc_id` varchar(11) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL DEFAULT NULL COMMENT '文档ID',
+                                       `doc_id` varchar(32) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL DEFAULT NULL COMMENT '文档ID',
                                       `content` text CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NOT NULL COMMENT '文档内容',
                                       `create_dept` varchar(255) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL DEFAULT NULL COMMENT '部门',
                                       `create_by` varchar(50) CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL DEFAULT NULL COMMENT '创建人',
@@ -1168,6 +1171,9 @@ CREATE TABLE `knowledge_fragment`  (
                                       `tenant_id` bigint NOT NULL DEFAULT 0 COMMENT '租户Id',
                                       `knowledge_id` bigint NULL DEFAULT NULL COMMENT '知识库ID',
                                       PRIMARY KEY (`id`) USING BTREE,
+                                       UNIQUE INDEX `uk_fid`(`fid`) USING BTREE,
+                                       INDEX `idx_doc_id`(`doc_id`) USING BTREE,
+                                       INDEX `idx_knowledge_id`(`knowledge_id`) USING BTREE,
                                       FULLTEXT INDEX `ft_content`(`content`) WITH PARSER `ngram`
 ) ENGINE = InnoDB AUTO_INCREMENT = 2033199209131880451 CHARACTER SET = utf8mb4 COLLATE = utf8mb4_0900_ai_ci COMMENT = '知识片段' ROW_FORMAT = DYNAMIC;

@@ -1206,7 +1212,9 @@ CREATE TABLE `knowledge_info`  (
                                   `enable_hybrid` tinyint(1) NULL DEFAULT 0 COMMENT '是否启用混合检索',
                                   `hybrid_alpha` double NULL DEFAULT 0.5 COMMENT '混合检索权重比例 (0.0=纯向量, 1.0=纯关键词)',
                                   `system_prompt` text CHARACTER SET utf8mb4 COLLATE utf8mb4_0900_ai_ci NULL COMMENT '系统提示词',
-                                   PRIMARY KEY (`id`) USING BTREE
+                                   PRIMARY KEY (`id`) USING BTREE,
+                                   INDEX `idx_tenant_user` (`tenant_id`, `user_id`) USING BTREE,
+                                   INDEX `idx_tenant_share` (`tenant_id`, `share`) USING BTREE
 ) ENGINE = InnoDB AUTO_INCREMENT = 2033198818050781187 CHARACTER SET = utf8mb4 COLLATE = utf8mb4_0900_ai_ci COMMENT = '知识库' ROW_FORMAT = DYNAMIC;

 -- ----------------------------
--- a/docs/script/sql/update/2026-07-20-knowledge-fragment-fid.sql
+++ b/docs/script/sql/update/2026-07-20-knowledge-fragment-fid.sql
@@ -0,0 +1,44 @@
+-- RAG metadata migration (MySQL 8). Safe to execute repeatedly.
+ALTER TABLE `knowledge_attach`
+    ADD COLUMN IF NOT EXISTS `file_hash` varchar(64) NULL DEFAULT NULL COMMENT '文件SHA-256摘要' AFTER `doc_id`,
+    MODIFY COLUMN `doc_id` varchar(32) NULL DEFAULT NULL COMMENT '文档ID';
+
+SET @add_file_hash = IF(EXISTS(
+    SELECT 1 FROM information_schema.statistics WHERE table_schema = DATABASE()
+    AND table_name = 'knowledge_attach' AND index_name = 'uk_knowledge_file_hash'),
+    'SELECT 1', 'ALTER TABLE `knowledge_attach` ADD UNIQUE INDEX `uk_knowledge_file_hash` (`knowledge_id`, `file_hash`)');
+PREPARE stmt FROM @add_file_hash; EXECUTE stmt; DEALLOCATE PREPARE stmt;
+
+ALTER TABLE `knowledge_fragment`
+    ADD COLUMN IF NOT EXISTS `fid` varchar(32) NULL DEFAULT NULL COMMENT '向量库片段ID' AFTER `id`,
+    MODIFY COLUMN `doc_id` varchar(32) NULL DEFAULT NULL COMMENT '文档ID';
+
+UPDATE `knowledge_fragment`
+SET `fid` = LOWER(MD5(CONCAT('knowledge_fragment:', `id`)))
+WHERE `fid` IS NULL OR `fid` = '';
+
+SET @drop_idx_fid = IF(EXISTS(
+    SELECT 1 FROM information_schema.statistics WHERE table_schema = DATABASE()
+    AND table_name = 'knowledge_fragment' AND index_name = 'idx_fid'),
+    'ALTER TABLE `knowledge_fragment` DROP INDEX `idx_fid`', 'SELECT 1');
+PREPARE stmt FROM @drop_idx_fid; EXECUTE stmt; DEALLOCATE PREPARE stmt;
+
+SET @add_uk_fid = IF(EXISTS(
+    SELECT 1 FROM information_schema.statistics WHERE table_schema = DATABASE()
+    AND table_name = 'knowledge_fragment' AND index_name = 'uk_fid'),
+    'SELECT 1', 'ALTER TABLE `knowledge_fragment` ADD UNIQUE INDEX `uk_fid` (`fid`)');
+PREPARE stmt FROM @add_uk_fid; EXECUTE stmt; DEALLOCATE PREPARE stmt;
+
+ALTER TABLE `knowledge_fragment` MODIFY COLUMN `fid` varchar(32) NOT NULL COMMENT '向量库片段ID';
+
+SET @add_tenant_user = IF(EXISTS(
+    SELECT 1 FROM information_schema.statistics WHERE table_schema = DATABASE()
+    AND table_name = 'knowledge_info' AND index_name = 'idx_tenant_user'),
+    'SELECT 1', 'ALTER TABLE `knowledge_info` ADD INDEX `idx_tenant_user` (`tenant_id`, `user_id`)');
+PREPARE stmt FROM @add_tenant_user; EXECUTE stmt; DEALLOCATE PREPARE stmt;
+
+SET @add_tenant_share = IF(EXISTS(
+    SELECT 1 FROM information_schema.statistics WHERE table_schema = DATABASE()
+    AND table_name = 'knowledge_info' AND index_name = 'idx_tenant_share'),
+    'SELECT 1', 'ALTER TABLE `knowledge_info` ADD INDEX `idx_tenant_share` (`tenant_id`, `share`)');
+PREPARE stmt FROM @add_tenant_share; EXECUTE stmt; DEALLOCATE PREPARE stmt;
--- a/pom.xml
+++ b/pom.xml
@@ -84,7 +84,7 @@
        <maven-surefire-plugin.version>3.5.3</maven-surefire-plugin.version>
        <flatten-maven-plugin.version>1.3.0</flatten-maven-plugin.version>
        <!-- 打包默认跳过测试 -->
-        <skipTests>true</skipTests>
+        <skipTests>false</skipTests>
    </properties>

    <profiles>
--- a/ruoyi-admin/src/main/resources/application-dev.yml
+++ b/ruoyi-admin/src/main/resources/application-dev.yml
@@ -60,7 +60,7 @@ spring:
          # rewriteBatchedStatements=true 批处理优化 大幅提升批量插入更新删除性能(对数据库有性能损耗 使用批量操作应考虑性能问题)
          url: jdbc:mysql://127.0.0.1:3306/ruoyi-ai?useUnicode=true&characterEncoding=utf8&zeroDateTimeBehavior=convertToNull&useSSL=true&serverTimezone=GMT%2B8&autoReconnect=true&rewriteBatchedStatements=true&allowPublicKeyRetrieval=true&nullCatalogMeansCurrent=true
          username: root
-          password: root
+          password: 123456
 #        agent:
 #          url: jdbc:mysql://127.0.0.1:3306/test?useUnicode=true&characterEncoding=utf8&zeroDateTimeBehavior=convertToNull&useSSL=true&serverTimezone=GMT%2B8&autoReconnect=true&rewriteBatchedStatements=true&allowPublicKeyRetrieval=true&nullCatalogMeansCurrent=true
 #          # url: jdbc:mysql://localhost:3306/agent_db
--- a/ruoyi-admin/src/main/resources/application.yml
+++ b/ruoyi-admin/src/main/resources/application.yml
@@ -301,11 +301,12 @@ warm-flow:
 vector-store:
  # 向量存储类型 可选(weaviate/milvus/qdrant)
  # 如需修改向量库类型，请修改此配置值!
-  type: milvus
+  # 注意：需与 docker-compose 实际部署的向量库保持一致（当前 compose 内置 weaviate，映射端口 28080）
+  type: weaviate
  # Weaviate配置
  weaviate:
    protocol: http
-    host: 127.0.0.1:6038
+    host: 127.0.0.1:28080
    classname: LocalKnowledge
  # Milvus配置
  milvus:
--- a/ruoyi-modules/ruoyi-aiflow/pom.xml
+++ b/ruoyi-modules/ruoyi-aiflow/pom.xml
@@ -29,6 +29,12 @@
            <artifactId>ruoyi-common-chat</artifactId>
        </dependency>

+        <!-- 复用聊天模块的知识库统一检索能力（向量/混合/重排） -->
+        <dependency>
+            <groupId>org.ruoyi</groupId>
+            <artifactId>ruoyi-chat</artifactId>
+        </dependency>
+
        <dependency>
            <groupId>org.ruoyi</groupId>
            <artifactId>ruoyi-common-web</artifactId>
--- a/ruoyi-modules/ruoyi-aiflow/src/main/java/org/ruoyi/workflow/workflow/node/knowledgeRetrieval/KnowledgeRetrievalNode.java
+++ b/ruoyi-modules/ruoyi-aiflow/src/main/java/org/ruoyi/workflow/workflow/node/knowledgeRetrieval/KnowledgeRetrievalNode.java
@@ -80,9 +80,10 @@ public class KnowledgeRetrievalNode extends AbstractWfNode {
        String retrievalResult;
        String mode = config.getRetrievalMode() != null ? config.getRetrievalMode().toLowerCase() : "vector";

-        // 目前只支持向量检索，图谱检索需要依赖graph模块
-        if ("graph".equals(mode) || "hybrid".equals(mode)) {
-            log.warn("Graph retrieval mode is not supported in workflow-api module, falling back to vector retrieval");
+        // 图谱检索需要依赖 graph 模块，暂不支持；vector/hybrid 由统一检索服务处理
+        if ("graph".equals(mode)) {
+            log.warn("Graph retrieval mode is not supported");
+            throw new UnsupportedOperationException("GraphRAG retrieval is not supported");
        }

        retrievalResult = retrieveFromVector(config, finalQuery);
@@ -203,18 +204,75 @@ public class KnowledgeRetrievalNode extends AbstractWfNode {
    }

    /**
-     * 从向量库检索
+     * 从向量库检索（复用聊天模块的统一检索服务：向量 + 可选混合检索 + 可选重排）
     */
    private String retrieveFromVector(KnowledgeRetrievalNodeConfig config, String query) {
        try {
-
-            // 获取知识库信息以获取embedding模型配置
            Long knowledgeId = Long.parseLong(config.getKnowledgeId());

-            // 合并结果
-            String mergedResult = "根据知识库id + query 查询知识库内容";
+            org.ruoyi.service.knowledge.IKnowledgeInfoService knowledgeInfoService =
+                SpringUtil.getBean(org.ruoyi.service.knowledge.IKnowledgeInfoService.class);
+            org.ruoyi.domain.vo.knowledge.KnowledgeInfoVo kb = knowledgeInfoService.queryById(knowledgeId);
+            if (kb == null) {
+                log.error("Knowledge base not found: {}", knowledgeId);
+                return "错误：知识库不存在, id=" + knowledgeId;
+            }

-            return mergedResult;
+            org.ruoyi.common.chat.service.chat.IChatModelService chatModelService =
+                SpringUtil.getBean(org.ruoyi.common.chat.service.chat.IChatModelService.class);
+            org.ruoyi.common.chat.domain.vo.chat.ChatModelVo embModel =
+                chatModelService.selectModelByName(kb.getEmbeddingModel());
+            if (embModel == null) {
+                log.error("Embedding model not found: {}", kb.getEmbeddingModel());
+                return "错误：知识库未配置有效的向量模型";
+            }
+
+            // 组装检索参数：节点配置优先，混合检索/重排继承知识库配置
+            org.ruoyi.domain.bo.vector.QueryVectorBo bo = new org.ruoyi.domain.bo.vector.QueryVectorBo();
+            bo.setQuery(query);
+            bo.setKid(String.valueOf(knowledgeId));
+            bo.setMaxResults(config.getTopK() != null ? config.getTopK() : kb.getRetrieveLimit());
+            bo.setSimilarityThreshold(config.getSimilarityThreshold() != null
+                ? config.getSimilarityThreshold() : kb.getSimilarityThreshold());
+            bo.setEmbeddingModelName(kb.getEmbeddingModel());
+            bo.setVectorModelName(kb.getVectorModel());
+            bo.setApiKey(embModel.getApiKey());
+            bo.setBaseUrl(embModel.getApiHost());
+
+            String mode = config.getRetrievalMode() != null ? config.getRetrievalMode().toLowerCase() : "vector";
+            boolean enableHybrid = "hybrid".equals(mode)
+                || (kb.getEnableHybrid() != null && kb.getEnableHybrid() == 1);
+            bo.setEnableHybrid(enableHybrid);
+            bo.setHybridAlpha(kb.getHybridAlpha());
+            bo.setEnableRerank(kb.getEnableRerank() != null && kb.getEnableRerank() == 1);
+            bo.setRerankModelName(kb.getRerankModel());
+            bo.setRerankTopN(kb.getRerankTopN());
+            bo.setRerankScoreThreshold(kb.getRerankScoreThreshold());
+
+            org.ruoyi.service.retrieval.KnowledgeRetrievalService retrievalService =
+                SpringUtil.getBean(org.ruoyi.service.retrieval.KnowledgeRetrievalService.class);
+            java.util.List<org.ruoyi.domain.vo.knowledge.KnowledgeRetrievalVo> results = retrievalService.retrieve(bo);
+            if (results == null || results.isEmpty()) {
+                log.info("Knowledge retrieval returned no results, kid={}, query={}", knowledgeId, query);
+                return "";
+            }
+
+            // 合并结果
+            boolean returnSource = config.getReturnSource() == null || config.getReturnSource();
+            StringBuilder sb = new StringBuilder();
+            for (int i = 0; i < results.size(); i++) {
+                org.ruoyi.domain.vo.knowledge.KnowledgeRetrievalVo vo = results.get(i);
+                sb.append(i + 1).append(". ").append(vo.getContent());
+                if (returnSource && StringUtils.isNotBlank(vo.getSourceName())) {
+                    sb.append("（来源: ").append(vo.getSourceName());
+                    if (vo.getScore() != null) {
+                        sb.append(String.format(", 相关度: %.3f", vo.getScore()));
+                    }
+                    sb.append("）");
+                }
+                sb.append("\n");
+            }
+            return sb.toString().trim();
        } catch (NumberFormatException e) {
            log.error("Invalid knowledge base ID format: {}", config.getKnowledgeId(), e);
            return "错误：知识库ID格式无效";
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/controller/knowledge/KnowledgeAttachController.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/controller/knowledge/KnowledgeAttachController.java
@@ -9,6 +9,7 @@ import cn.dev33.satoken.annotation.SaCheckPermission;
 import org.ruoyi.domain.bo.knowledge.KnowledgeAttachBo;
 import org.ruoyi.domain.bo.knowledge.KnowledgeInfoUploadBo;
 import org.ruoyi.domain.vo.knowledge.KnowledgeAttachVo;
+import org.ruoyi.domain.vo.knowledge.KnowledgeReparseVo;
 import org.ruoyi.service.knowledge.IKnowledgeAttachService;
 import org.springframework.web.bind.annotation.*;
 import org.springframework.validation.annotation.Validated;
@@ -106,7 +107,10 @@ public class KnowledgeAttachController extends BaseController {

    /**
     * 上传知识库附件
+     * 注意：multipart 上传不能加 @RepeatSubmit（其参数序列化不支持 MultipartFile）
     */
+    @SaCheckPermission("system:attach:add")
+    @Log(title = "知识库附件", businessType = BusinessType.INSERT)
    @PostMapping(value = "/upload")
    public R<String> upload(KnowledgeInfoUploadBo bo){
        knowledgeAttachService.upload(bo);
@@ -118,9 +122,20 @@ public class KnowledgeAttachController extends BaseController {
     *
     * @param id 附件ID
     */
+    @SaCheckPermission("system:attach:edit")
+    @Log(title = "知识库附件", businessType = BusinessType.UPDATE)
    @PostMapping("/parse/{id}")
+    @RepeatSubmit()
    public R<Void> parse(@PathVariable Long id) {
        knowledgeAttachService.parse(id);
        return R.ok();
    }
+
+    @SaCheckPermission("system:attach:edit")
+    @Log(title = "知识库附件批量重新解析", businessType = BusinessType.UPDATE)
+    @PostMapping("/reparse/knowledge/{knowledgeId}")
+    @RepeatSubmit()
+    public R<KnowledgeReparseVo> reparseKnowledge(@PathVariable Long knowledgeId) {
+        return R.ok(knowledgeAttachService.reparseKnowledge(knowledgeId));
+    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/controller/knowledge/KnowledgeFragmentController.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/controller/knowledge/KnowledgeFragmentController.java
@@ -107,7 +107,9 @@ public class KnowledgeFragmentController extends BaseController {
    /**
     * 检索测试
     */
+    @SaCheckPermission("system:fragment:list")
    @PostMapping("/retrieval")
+    @RepeatSubmit()
    public R<List<KnowledgeRetrievalVo>> retrieval(@RequestBody KnowledgeFragmentBo bo) {
        return R.ok(knowledgeFragmentService.retrieval(bo));
    }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/domain/entity/knowledge/KnowledgeAttach.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/domain/entity/knowledge/KnowledgeAttach.java
@@ -37,6 +37,9 @@ public class KnowledgeAttach extends BaseEntity {
     */
    private String docId;

+    /** SHA-256 content digest used for upload idempotency. */
+    private String fileHash;
+
    /**
     * 附件名称
     */
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/domain/entity/knowledge/KnowledgeFragment.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/domain/entity/knowledge/KnowledgeFragment.java
@@ -27,6 +27,11 @@ public class KnowledgeFragment extends BaseEntity {
    @TableId(value = "id")
    private Long id;

+    /**
+     * 向量库片段ID（与向量库中的 fid 元数据对应，用于向量定位与混合检索融合）
+     */
+    private String fid;
+
    /**
     * 文档ID-用于关联文本块信息
     */
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/domain/vo/knowledge/KnowledgeFragmentVo.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/domain/vo/knowledge/KnowledgeFragmentVo.java
@@ -30,6 +30,11 @@ public class KnowledgeFragmentVo implements Serializable {
    @ExcelProperty(value = "主键")
    private Long id;

+    /**
+     * 向量库片段ID
+     */
+    private String fid;
+
    /**
     * 文档ID-用于关联文本块信息
     */
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/domain/vo/knowledge/KnowledgeReparseVo.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/domain/vo/knowledge/KnowledgeReparseVo.java
@@ -0,0 +1,4 @@
+package org.ruoyi.domain.vo.knowledge;
+
+public record KnowledgeReparseVo(int submitted, int skipped, int total) {
+}
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/factory/EmbeddingModelFactory.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/factory/EmbeddingModelFactory.java
@@ -77,13 +77,22 @@ public class EmbeddingModelFactory {

    /**
     * 刷新模型缓存
-     * 根据给定的嵌入模型ID从缓存中移除对应的模型
+     * 根据给定的嵌入模型ID解析模型名称后，从缓存中移除对应的模型
     *
     * @param embeddingModelId 嵌入模型的唯一标识ID
     */
    public void refreshModel(Long embeddingModelId) {
-        // 从模型缓存中移除指定ID的模型
-        modelCache.remove(embeddingModelId);
+        ChatModelVo modelConfig = chatModelService.queryById(embeddingModelId);
+        if (modelConfig != null) {
+            modelCache.remove(modelConfig.getModelName());
+        }
+    }
+
+    /**
+     * 按模型名称刷新缓存
+     */
+    public void refreshModelByName(String embeddingModelName) {
+        modelCache.remove(embeddingModelName);
    }

    /**
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/factory/RerankModelFactory.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/factory/RerankModelFactory.java
@@ -55,12 +55,22 @@ public class RerankModelFactory {

    /**
     * 刷新模型缓存
-     * 根据给定的模型ID从缓存中移除对应的模型
+     * 根据给定的模型ID解析模型名称后，从缓存中移除对应的模型
     *
     * @param modelId 模型的唯一标识ID
     */
    public void refreshModel(Long modelId) {
-        modelCache.remove(modelId);
+        ChatModelVo modelConfig = chatModelService.queryById(modelId);
+        if (modelConfig != null) {
+            modelCache.remove(modelConfig.getModelName());
+        }
+    }
+
+    /**
+     * 按模型名称刷新缓存
+     */
+    public void refreshModelByName(String modelName) {
+        modelCache.remove(modelName);
    }

    /**
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/factory/ResourceLoaderFactory.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/factory/ResourceLoaderFactory.java
@@ -17,7 +17,7 @@ public class ResourceLoaderFactory {
    private final ExcelTextSplitter excelTextSplitter;

    public ResourceLoader getLoaderByFileType(String fileType) {
-        fileType = StringUtils.removeStart(fileType, ".");
+        fileType = StringUtils.lowerCase(StringUtils.removeStart(StringUtils.trim(fileType), "."));
        if (FileTypeConstants.isTextFile(fileType)) {
            return new TextFileLoader(characterTextSplitter);
        } else if (FileTypeConstants.isWord(fileType)) {
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/factory/VectorStoreStrategyFactory.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/factory/VectorStoreStrategyFactory.java
@@ -10,6 +10,7 @@ import org.ruoyi.service.vector.impl.MilvusVectorStoreStrategy;
 import org.ruoyi.service.vector.impl.QdrantVectorStoreStrategy;
 import org.ruoyi.service.vector.impl.WeaviateVectorStoreStrategy;
 import org.springframework.stereotype.Component;
+import org.ruoyi.common.core.exception.ServiceException;

 import java.util.HashMap;
 import java.util.Map;
@@ -45,14 +46,20 @@ public class VectorStoreStrategyFactory {
     * 获取当前配置的向量库策略
     */
    public VectorStoreService getStrategy() {
-        String vectorStoreType = vectorStoreProperties.getType();
+        return getStrategy(null);
+    }
+
+    public VectorStoreService getStrategy(String requestedType) {
+        String vectorStoreType = requestedType;
+        if (vectorStoreType == null || vectorStoreType.trim().isEmpty()) {
+            vectorStoreType = vectorStoreProperties.getType();
+        }
        if (vectorStoreType == null || vectorStoreType.trim().isEmpty()) {
            vectorStoreType = "weaviate"; // 默认使用weaviate
        }
        VectorStoreService strategy = strategies.get(vectorStoreType.toLowerCase());
        if (strategy == null) {
-            log.warn("未找到向量库策略: {}, 使用默认策略: weaviate", vectorStoreType);
-            strategy = strategies.get("weaviate");
+            throw new ServiceException("不支持的向量库类型: " + vectorStoreType);
        }
        log.debug("使用向量库策略: {}", vectorStoreType);
        return strategy;
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/mapper/knowledge/KnowledgeAttachMapper.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/mapper/knowledge/KnowledgeAttachMapper.java
@@ -6,6 +6,8 @@ import org.apache.ibatis.annotations.Select;
 import org.ruoyi.domain.entity.knowledge.KnowledgeAttach;
 import org.ruoyi.domain.vo.knowledge.KnowledgeAttachVo;
 import org.ruoyi.common.mybatis.core.mapper.BaseMapperPlus;
+import java.util.List;
+import java.util.Map;

 /**
 * 知识库附件Mapper接口
@@ -21,4 +23,10 @@ public interface KnowledgeAttachMapper extends BaseMapperPlus<KnowledgeAttach, K
     */
    @Select("SELECT COUNT(*) FROM knowledge_attach WHERE knowledge_id = #{knowledgeId}")
    int countByKnowledgeId(@Param("knowledgeId") Long knowledgeId);
+
+    @Select("<script>SELECT knowledge_id AS knowledgeId, COUNT(*) AS documentCount " +
+            "FROM knowledge_attach WHERE knowledge_id IN " +
+            "<foreach collection='knowledgeIds' item='id' open='(' separator=',' close=')'>#{id}</foreach> " +
+            "GROUP BY knowledge_id</script>")
+    List<Map<String, Object>> countByKnowledgeIds(@Param("knowledgeIds") List<Long> knowledgeIds);
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/mapper/knowledge/KnowledgeFragmentMapper.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/mapper/knowledge/KnowledgeFragmentMapper.java
@@ -34,7 +34,7 @@ public interface KnowledgeFragmentMapper extends BaseMapperPlus<KnowledgeFragmen
            "</script>")
    List<DocFragmentCountVo> selectFragmentCountByDocIds(@Param("docIds") List<String> docIds);
    @Select("<script>" +
-            "SELECT id, doc_id AS docId, content, idx, knowledge_id AS knowledgeId " +
+            "SELECT id, fid, doc_id AS docId, content, idx, knowledge_id AS knowledgeId " +
            "FROM knowledge_fragment " +
            "WHERE knowledge_id = #{knowledgeId} " +
            "AND MATCH (content) AGAINST (#{query} IN NATURAL LANGUAGE MODE) " +
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/chat/impl/ChatServiceFacade.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/chat/impl/ChatServiceFacade.java
@@ -77,6 +77,7 @@ import org.springframework.web.servlet.mvc.method.annotation.SseEmitter;
 import java.util.ArrayList;
 import java.util.List;
 import java.util.Map;
+import java.util.LinkedHashMap;
 import java.util.concurrent.CompletableFuture;
 import java.util.concurrent.ConcurrentHashMap;

@@ -161,7 +162,8 @@ public class ChatServiceFacade implements IChatService {
            throw new IllegalArgumentException("模型不存在: " + chatRequest.getModel());
        }

-        // 2. 构建上下文消息列表
+        // 2. 构建上下文消息列表（系统提示词 + 历史消息 + 当前用户消息）
+        //    注意：RAG 检索增强统一在 handleAgentChat 中执行一次，此处不再重复检索
        List<ChatMessage> contextMessages = buildContextMessages(chatRequest, agentVo);

        chatRequest.setEmitter(emitter);
@@ -282,12 +284,20 @@ public class ChatServiceFacade implements IChatService {
            .responseStrategy(SupervisorResponseStrategy.SUMMARY);
        SupervisorAgent supervisor = supervisorBuilder.build();

-        // 知识库增强：智能体绑定了知识库时，对 supervisor 输入做一次 RAG 增强
+        // 知识库增强：智能体绑定了知识库时，对 supervisor 输入做一次 RAG 增强（全程唯一一次检索）
        String augmentedInput = augmentAgentInput(chatRequest, agentVo);
-        // 智能体自定义系统提示词：supervisor builder 不支持 systemMessage，前置到输入
-        String prompt = (agentVo != null && StringUtils.isNotBlank(agentVo.getSystemPrompt()))
-            ? agentVo.getSystemPrompt() + "\n\n" + augmentedInput
-            : augmentedInput;
+        // 组装最终 prompt：系统提示词 → 多轮历史 → RAG 增强后的当前提问
+        StringBuilder promptBuilder = new StringBuilder();
+        if (agentVo != null && StringUtils.isNotBlank(agentVo.getSystemPrompt())) {
+            promptBuilder.append(agentVo.getSystemPrompt()).append("\n\n");
+        }
+        String historyText = formatHistoryMessages(chatRequest.getContextMessages(), chatRequest.getContent());
+        if (StringUtils.isNotBlank(historyText)) {
+            promptBuilder.append("以下是本次会话的历史对话，请结合上下文理解用户最新提问：\n")
+                .append(historyText).append("\n\n");
+        }
+        promptBuilder.append(augmentedInput);
+        String prompt = promptBuilder.toString();

        String tokenValue = chatRequest.getTokenValue();

@@ -316,8 +326,9 @@ public class ChatServiceFacade implements IChatService {
     * 兜底 MCP 工具装配（无智能体时使用，保留原有 3 个硬编码客户端逻辑）
     */
    private ToolProvider buildDefaultMcpToolProvider(Long userId) {
+        String npxCommand = resolveNpxCommand();
        McpTransport playwrightTransport = new StdioMcpTransport.Builder()
-            .command(List.of("C:\\Program Files\\nodejs\\npx.cmd", "-y", "@playwright/mcp@latest"))
+            .command(List.of(npxCommand, "-y", "@playwright/mcp@latest"))
            .logEvents(true)
            .build();
        McpClient playwrightMcpClient = new DefaultMcpClient.Builder()
@@ -327,7 +338,7 @@ public class ChatServiceFacade implements IChatService {

        String userDir = System.getProperty("user.dir");
        McpTransport filesystemTransport = new StdioMcpTransport.Builder()
-            .command(List.of("C:\\Program Files\\nodejs\\npx.cmd", "-y",
+            .command(List.of(npxCommand, "-y",
                "@modelcontextprotocol/server-filesystem", userDir))
            .logEvents(true)
            .build();
@@ -341,6 +352,14 @@ public class ChatServiceFacade implements IChatService {
            .build();
    }

+    private String resolveNpxCommand() {
+        String configured = System.getProperty("mcp.npx.command");
+        if (StringUtils.isNotBlank(configured)) return configured;
+        String fromEnv = System.getenv("MCP_NPX_COMMAND");
+        if (StringUtils.isNotBlank(fromEnv)) return fromEnv;
+        return System.getProperty("os.name", "").toLowerCase().contains("win") ? "npx.cmd" : "npx";
+    }
+
    /**
     * 装配磁盘 ShellSkills：智能体勾选了技能名时按名过滤，否则加载全部。
     * 无 skills 时返回 null（调用方据此跳过 SkillsAgent 的 toolProvider 注入）
@@ -476,27 +495,7 @@ public class ChatServiceFacade implements IChatService {
            messages.add(SystemMessage.from(agentVo.getSystemPrompt()));
        }

-        // 1. 初始化当前用户消息
-        UserMessage userMessage = UserMessage.userMessage(chatRequest.getContent());
-
-        // 2. 知识库检索增强 (RAG)：智能体的 knowledgeIds 优先，回退到请求的 knowledgeId
-        List<Long> knowledgeIds = collectKnowledgeIds(chatRequest, agentVo);
-        if (knowledgeIds != null && !knowledgeIds.isEmpty()) {
-            RetrievalAugmentor augmentor = buildMultiKnowledgeAugmentor(knowledgeIds);
-            if (augmentor != null) {
-                log.info("执行多知识库 RAG 流程: kids={}", knowledgeIds);
-                Metadata metadata = Metadata.from(userMessage, chatRequest.getSessionId(), new ArrayList<>());
-                AugmentationRequest augmentationRequest = new AugmentationRequest(userMessage, metadata);
-                AugmentationResult result = augmentor.augment(augmentationRequest);
-                ChatMessage augmented = result.chatMessage();
-                if (augmented instanceof UserMessage) {
-                    userMessage = (UserMessage) augmented;
-                    log.debug("RAG 增强完成，UserMessage 已注入背景知识");
-                }
-            }
-        }
-
-        // 3. 从数据库查询历史对话消息（放在前面）
+        // 1. 从数据库查询历史对话消息（放在前面）
        if (chatRequest.getSessionId() != null) {
            MessageWindowChatMemory memory = createChatMemory(chatRequest.getSessionId());
            if (memory != null) {
@@ -508,12 +507,37 @@ public class ChatServiceFacade implements IChatService {
            }
        }

-        // 4. 添加经过增强的用户消息（放在最后）
-        messages.add(userMessage);
+        // 2. 添加当前用户消息（放在最后；RAG 增强在 handleAgentChat 中统一执行，避免重复检索）
+        messages.add(UserMessage.userMessage(chatRequest.getContent()));

        return messages;
    }

+    /**
+     * 将上下文消息格式化为多轮对话文本（供只接受 String 输入的 Supervisor 使用）。
+     * 跳过 SystemMessage（系统提示词单独前置）与最后一条当前用户消息（单独做 RAG 增强后拼接）。
+     */
+    private String formatHistoryMessages(List<ChatMessage> contextMessages, String currentContent) {
+        if (contextMessages == null || contextMessages.isEmpty()) {
+            return "";
+        }
+        StringBuilder sb = new StringBuilder();
+        int limit = contextMessages.size();
+        // 最后一条是当前用户消息，不纳入历史（避免与增强后的输入重复）
+        if (limit > 0 && contextMessages.get(limit - 1) instanceof UserMessage) {
+            limit--;
+        }
+        for (int i = 0; i < limit; i++) {
+            ChatMessage msg = contextMessages.get(i);
+            if (msg instanceof UserMessage userMsg) {
+                sb.append("用户: ").append(userMsg.singleText()).append("\n");
+            } else if (msg instanceof AiMessage aiMsg) {
+                sb.append("助手: ").append(aiMsg.text()).append("\n");
+            }
+        }
+        return sb.toString().trim();
+    }
+
    /**
     * 汇总本次对话要检索的知识库ID列表：智能体绑定的 knowledgeIds 优先，回退到请求的 knowledgeId
     */
@@ -580,18 +604,35 @@ public class ChatServiceFacade implements IChatService {

        @Override
        public List<Content> retrieve(Query query) {
-            List<Content> all = new ArrayList<>();
-            for (ContentRetriever r : delegates) {
-                try {
-                    List<Content> part = r.retrieve(query);
-                    if (part != null) {
-                        all.addAll(part);
-                    }
-                } catch (Exception e) {
-                    log.warn("复合检索子检索器异常: {}", e.getMessage());
+            List<CompletableFuture<List<Content>>> futures = delegates.stream()
+                    .map(r -> CompletableFuture.supplyAsync(() -> {
+                        try {
+                            List<Content> part = r.retrieve(query);
+                            return part == null ? List.<Content>of() : part;
+                        } catch (Exception e) {
+                            log.warn("复合检索子检索器异常: {}", e.getMessage());
+                            return List.<Content>of();
+                        }
+                    })).toList();
+            Map<String, Content> unique = new LinkedHashMap<>();
+            for (CompletableFuture<List<Content>> future : futures) {
+                for (Content content : future.join()) {
+                    String key = content.textSegment().metadata().getString("kid") + "|"
+                            + content.textSegment().metadata().getString("docId") + "|"
+                            + content.textSegment().metadata().getString("fid");
+                    if (key.endsWith("null|null|null")) key = content.textSegment().text();
+                    unique.putIfAbsent(key, content);
                }
            }
-            return all;
+            List<Content> bounded = new ArrayList<>();
+            int chars = 0;
+            for (Content content : unique.values()) {
+                int next = content.textSegment().text().length();
+                if (bounded.size() >= 20 || chars + next > 24000) break;
+                bounded.add(content);
+                chars += next;
+            }
+            return bounded;
        }
    }

--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/AliBaiLianBaseEmbedProvider.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/AliBaiLianBaseEmbedProvider.java
@@ -19,6 +19,7 @@ import java.util.Set;
 * @Description: 阿里百炼基础嵌入模型（兼容openai）
 */
@Component("alibailian")
+@org.springframework.context.annotation.Scope("prototype")
 public class AliBaiLianBaseEmbedProvider extends OpenAiEmbeddingProvider {

    private ChatModelVo chatModelVo;
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/AliBaiLianMultiEmbeddingProvider.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/AliBaiLianMultiEmbeddingProvider.java
@@ -24,6 +24,7 @@ import java.util.concurrent.TimeUnit;
 * 实现了MultiModalEmbedModelService接口，提供文本、图像和视频的嵌入向量生成服务
 */
@Component("bailianMultiModel")
+@org.springframework.context.annotation.Scope("prototype")
@Slf4j
 public class AliBaiLianMultiEmbeddingProvider implements MultiModalEmbedModelService {
    private final OkHttpClient okHttpClient;
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/MinimaxEmbeddingProvider.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/MinimaxEmbeddingProvider.java
@@ -12,6 +12,7 @@ import org.springframework.stereotype.Component;
 * @date 2026/3/21
 */
@Component("minimax")
+@org.springframework.context.annotation.Scope("prototype")
 public class MinimaxEmbeddingProvider extends OpenAiEmbeddingProvider {

 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/OllamaEmbeddingProvider.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/OllamaEmbeddingProvider.java
@@ -19,6 +19,7 @@ import java.util.Set;
 * @Description: Ollama嵌入模型
 */
@Component("ollama")
+@org.springframework.context.annotation.Scope("prototype")
 public class OllamaEmbeddingProvider implements BaseEmbedModelService {
    private ChatModelVo chatModelVo;

--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/OpenAiEmbeddingProvider.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/OpenAiEmbeddingProvider.java
@@ -19,6 +19,7 @@ import java.util.Set;
 * @Description: OpenAi嵌入模型
 */
@Component("openai")
+@org.springframework.context.annotation.Scope("prototype")
 public class OpenAiEmbeddingProvider implements BaseEmbedModelService {
    protected ChatModelVo chatModelVo;

--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/SiliconFlowEmbeddingProvider.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/SiliconFlowEmbeddingProvider.java
@@ -9,6 +9,7 @@ import org.springframework.stereotype.Component;
 * @Description: 硅基流动（兼容 OpenAi）
 */
@Component("siliconflow")
+@org.springframework.context.annotation.Scope("prototype")
 public class SiliconFlowEmbeddingProvider extends OpenAiEmbeddingProvider {

 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/ZhipuAiEmbeddingProvider.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/embed/impl/ZhipuAiEmbeddingProvider.java
@@ -21,6 +21,7 @@ import java.util.Set;
 * @Description: 智谱AI嵌入模型
 */
@Component("zhipu")
+@org.springframework.context.annotation.Scope("prototype")
 public class ZhipuAiEmbeddingProvider implements BaseEmbedModelService {
    protected ChatModelVo chatModelVo;

--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/DocumentSplitConfig.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/DocumentSplitConfig.java
@@ -0,0 +1,20 @@
+package org.ruoyi.service.knowledge;
+
+import org.ruoyi.common.core.exception.ServiceException;
+
+/** Immutable snapshot of the split settings used for one parse operation. */
+public record DocumentSplitConfig(String separator, int blockSize, int overlap, String fileType) {
+
+    public static final int DEFAULT_BLOCK_SIZE = 1000;
+    public static final int DEFAULT_OVERLAP = 50;
+
+    public DocumentSplitConfig {
+        if (blockSize <= 0) {
+            throw new ServiceException("文本块大小必须大于0");
+        }
+        if (overlap < 0 || overlap >= blockSize) {
+            throw new ServiceException("重叠字符数必须大于等于0且小于文本块大小");
+        }
+        fileType = fileType == null ? "" : fileType.strip().replaceFirst("^\\.", "").toLowerCase();
+    }
+}
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/IKnowledgeAttachService.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/IKnowledgeAttachService.java
@@ -5,6 +5,7 @@ import org.ruoyi.common.mybatis.core.page.PageQuery;
 import org.ruoyi.domain.bo.knowledge.KnowledgeAttachBo;
 import org.ruoyi.domain.bo.knowledge.KnowledgeInfoUploadBo;
 import org.ruoyi.domain.vo.knowledge.KnowledgeAttachVo;
+import org.ruoyi.domain.vo.knowledge.KnowledgeReparseVo;

 import java.util.Collection;
 import java.util.List;
@@ -79,4 +80,6 @@ public interface IKnowledgeAttachService {
     * @param id 附件ID
     */
    void parse(Long id);
+
+    KnowledgeReparseVo reparseKnowledge(Long knowledgeId);
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/ResourceLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/ResourceLoader.java
@@ -10,5 +10,5 @@ public interface ResourceLoader {

    String getContent(InputStream inputStream);

-    List<String> getChunkList(String content, String kid);
+    List<String> getChunkList(String content, DocumentSplitConfig config);
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/TextSplitter.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/TextSplitter.java
@@ -11,8 +11,8 @@ public interface TextSplitter {
     * 文本切分
     *
     * @param content 文本内容
-     * @param kid     知识库id
+     * @param config  本次解析的分片配置快照
     * @return 切分后的文本列表
     */
-    List<String> split(String content, String kid);
+    List<String> split(String content, DocumentSplitConfig config);
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/KnowledgeAttachServiceImpl.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/KnowledgeAttachServiceImpl.java
@@ -2,6 +2,7 @@ package org.ruoyi.service.knowledge.impl;

 import cn.hutool.core.collection.CollUtil;
 import cn.hutool.core.util.RandomUtil;
+import cn.hutool.crypto.digest.DigestUtil;
 import com.baomidou.mybatisplus.core.conditions.query.LambdaQueryWrapper;
 import com.baomidou.mybatisplus.core.toolkit.Wrappers;
 import com.baomidou.mybatisplus.extension.plugins.pagination.Page;
@@ -11,6 +12,7 @@ import org.ruoyi.common.chat.domain.vo.chat.ChatModelVo;
 import org.ruoyi.common.chat.service.chat.IChatModelService;
 import org.ruoyi.enums.KnowledgeAttachStatus;
 import org.ruoyi.common.core.domain.dto.OssDTO;
+import org.ruoyi.common.core.exception.ServiceException;
 import org.ruoyi.common.core.service.OssService;
 import org.ruoyi.common.core.utils.MapstructUtils;
 import org.ruoyi.common.core.utils.SpringUtils;
@@ -25,13 +27,16 @@ import org.ruoyi.domain.entity.knowledge.KnowledgeFragment;
 import org.ruoyi.domain.vo.knowledge.DocFragmentCountVo;
 import org.ruoyi.domain.vo.knowledge.KnowledgeAttachVo;
 import org.ruoyi.domain.vo.knowledge.KnowledgeInfoVo;
+import org.ruoyi.domain.vo.knowledge.KnowledgeReparseVo;
 import org.ruoyi.factory.ResourceLoaderFactory;
 import org.ruoyi.mapper.knowledge.KnowledgeAttachMapper;
 import org.ruoyi.mapper.knowledge.KnowledgeFragmentMapper;
 import org.ruoyi.service.knowledge.IKnowledgeAttachService;
 import org.ruoyi.service.knowledge.IKnowledgeInfoService;
 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;
 import org.ruoyi.service.vector.VectorStoreService;
+import org.ruoyi.service.retrieval.KnowledgeRetrievalService;
 import org.springframework.scheduling.annotation.Async;
 import org.springframework.stereotype.Service;
 import org.springframework.web.multipart.MultipartFile;
@@ -60,6 +65,7 @@ public class KnowledgeAttachServiceImpl implements IKnowledgeAttachService {
    private final ResourceLoaderFactory resourceLoaderFactory;
    private final VectorStoreService vectorStoreService;
    private final OssService ossService;
+    private final KnowledgeRetrievalService knowledgeRetrievalService;

    @Override
    public KnowledgeAttachVo queryById(Long id) {
@@ -126,18 +132,44 @@ public class KnowledgeAttachServiceImpl implements IKnowledgeAttachService {

    @Override
    public Boolean deleteWithValidByIds(Collection<Long> ids, Boolean isValid) {
+        // 删除附件前，同步清理其片段记录与向量库中的向量
+        List<KnowledgeAttach> attaches = baseMapper.selectByIds(ids);
+        for (KnowledgeAttach attach : attaches) {
+            String docId = attach.getDocId();
+            String kid = String.valueOf(attach.getKnowledgeId());
+            vectorStoreService.removeByDocId(docId, kid);
+            knowledgeFragmentMapper.delete(
+                Wrappers.<KnowledgeFragment>lambdaQuery().eq(KnowledgeFragment::getDocId, docId));
+            if (attach.getOssId() != null) {
+                ossService.deleteFile(attach.getOssId());
+            }
+            knowledgeRetrievalService.invalidateKnowledge(kid);
+        }
        return baseMapper.deleteByIds(ids) > 0;
    }

    @Override
    public void upload(KnowledgeInfoUploadBo bo) {
        MultipartFile file = bo.getFile();
+        final String fileHash;
+        try (InputStream input = file.getInputStream()) {
+            fileHash = DigestUtil.sha256Hex(input);
+        } catch (Exception e) {
+            throw new ServiceException("计算文件摘要失败", e);
+        }
+        boolean duplicate = baseMapper.exists(Wrappers.<KnowledgeAttach>lambdaQuery()
+            .eq(KnowledgeAttach::getKnowledgeId, bo.getKnowledgeId())
+            .eq(KnowledgeAttach::getFileHash, fileHash));
+        if (duplicate) {
+            throw new ServiceException("该文件已上传，请勿重复提交");
+        }
        OssDTO ossDTO = ossService.uploadFile(file);

        KnowledgeAttach knowledgeAttach = new KnowledgeAttach();
        knowledgeAttach.setKnowledgeId(bo.getKnowledgeId());
        knowledgeAttach.setOssId(ossDTO.getOssId());
        knowledgeAttach.setDocId(RandomUtil.randomString(10));
+        knowledgeAttach.setFileHash(fileHash);
        knowledgeAttach.setName(ossDTO.getOriginalName());
        knowledgeAttach.setType(ossDTO.getFileSuffix());
        knowledgeAttach.setStatus(KnowledgeAttachStatus.WAITING.getCode()); // 待解析
@@ -154,10 +186,17 @@ public class KnowledgeAttachServiceImpl implements IKnowledgeAttachService {
    @Override
    public void parse(Long id) {
        KnowledgeAttach attach = baseMapper.selectById(id);
-        if (attach == null || (!KnowledgeAttachStatus.WAITING.getCode().equals(attach.getStatus()) && !KnowledgeAttachStatus.FAILED.getCode().equals(attach.getStatus()))) {
+        if (attach == null || KnowledgeAttachStatus.PARSING.getCode().equals(attach.getStatus())) {
            return;
        }

+        int claimed = baseMapper.update(null, Wrappers.<KnowledgeAttach>lambdaUpdate()
+            .set(KnowledgeAttach::getStatus, KnowledgeAttachStatus.PARSING.getCode())
+            .set(KnowledgeAttach::getRemark, null)
+            .eq(KnowledgeAttach::getId, id)
+            .ne(KnowledgeAttach::getStatus, KnowledgeAttachStatus.PARSING.getCode()));
+        if (claimed == 0) return;
+
        try {
            attach.setStatus(KnowledgeAttachStatus.PARSING.getCode()); // 解析中
            baseMapper.updateById(attach);
@@ -166,6 +205,18 @@ public class KnowledgeAttachServiceImpl implements IKnowledgeAttachService {

            Long knowledgeId = attach.getKnowledgeId();
            String docId = attach.getDocId();
+            KnowledgeInfoVo knowledgeInfoVo = knowledgeInfoService.queryById(knowledgeId);
+            if (knowledgeInfoVo == null) {
+                throw new ServiceException("知识库不存在: " + knowledgeId);
+            }
+            int blockSize = knowledgeInfoVo.getTextBlockSize() == null
+                ? DocumentSplitConfig.DEFAULT_BLOCK_SIZE : knowledgeInfoVo.getTextBlockSize().intValue();
+            int overlap = knowledgeInfoVo.getOverlapChar() == null
+                ? DocumentSplitConfig.DEFAULT_OVERLAP : knowledgeInfoVo.getOverlapChar().intValue();
+            DocumentSplitConfig splitConfig = new DocumentSplitConfig(
+                knowledgeInfoVo.getSeparator(), blockSize, overlap, attach.getType());
+            List<KnowledgeFragment> oldFragments = knowledgeFragmentMapper.selectList(
+                    Wrappers.<KnowledgeFragment>lambdaQuery().eq(KnowledgeFragment::getDocId, docId));

            // 获取文件信息并下载
            List<OssDTO> ossDTOs = ossService.selectByIds(String.valueOf(attach.getOssId()));
@@ -178,28 +229,27 @@ public class KnowledgeAttachServiceImpl implements IKnowledgeAttachService {
            try (InputStream inputStream = new URL(ossDTO.getUrl()).openStream()) {
                content = resourceLoader.getContent(inputStream);
            }
-            List<String> chunkList = resourceLoader.getChunkList(content, String.valueOf(knowledgeId));
+            List<String> chunkList = resourceLoader.getChunkList(content, splitConfig);

-            List<String> fids = new ArrayList<>();
-            List<KnowledgeFragment> knowledgeFragmentList = new ArrayList<>();
-            if (CollUtil.isNotEmpty(chunkList)) {
-                for (int i = 0; i < chunkList.size(); i++) {
-                    String fid = RandomUtil.randomString(10);
-                    fids.add(fid);
-                    KnowledgeFragment knowledgeFragment = new KnowledgeFragment();
-                    knowledgeFragment.setKnowledgeId(knowledgeId);
-                    knowledgeFragment.setDocId(docId);
-                    knowledgeFragment.setIdx(i);
-                    knowledgeFragment.setContent(chunkList.get(i));
-                    knowledgeFragment.setCreateTime(new Date());
-                    knowledgeFragmentList.add(knowledgeFragment);
-                }
-                knowledgeFragmentMapper.delete(Wrappers.<KnowledgeFragment>lambdaQuery().eq(KnowledgeFragment::getDocId, docId));
-                knowledgeFragmentMapper.insertBatch(knowledgeFragmentList);
-                log.info("文档切片并入库完成，共计 {} 个片段。id: {}", chunkList.size(), id);
+            if (CollUtil.isEmpty(chunkList)) {
+                throw new RuntimeException("文档分片结果为空，请检查文档内容或分片器是否支持该文件类型");
            }

-            KnowledgeInfoVo knowledgeInfoVo = knowledgeInfoService.queryById(knowledgeId);
+            // 重新解析前先清理旧的向量数据，避免向量重复累积
+            List<String> fids = new ArrayList<>();
+            List<KnowledgeFragment> knowledgeFragmentList = new ArrayList<>();
+            for (int i = 0; i < chunkList.size(); i++) {
+                String fid = RandomUtil.randomString(10);
+                fids.add(fid);
+                KnowledgeFragment knowledgeFragment = new KnowledgeFragment();
+                knowledgeFragment.setKnowledgeId(knowledgeId);
+                knowledgeFragment.setDocId(docId);
+                knowledgeFragment.setFid(fid);
+                knowledgeFragment.setIdx(i);
+                knowledgeFragment.setContent(chunkList.get(i));
+                knowledgeFragment.setCreateTime(new Date());
+                knowledgeFragmentList.add(knowledgeFragment);
+            }
            ChatModelVo chatModelVo = chatModelService.selectModelByName(knowledgeInfoVo.getEmbeddingModel());

            StoreEmbeddingBo storeEmbeddingBo = new StoreEmbeddingBo();
@@ -211,7 +261,27 @@ public class KnowledgeAttachServiceImpl implements IKnowledgeAttachService {
            storeEmbeddingBo.setEmbeddingModelName(knowledgeInfoVo.getEmbeddingModel());
            storeEmbeddingBo.setApiKey(chatModelVo.getApiKey());
            storeEmbeddingBo.setBaseUrl(chatModelVo.getApiHost());
-            vectorStoreService.storeEmbeddings(storeEmbeddingBo);
+            try {
+                vectorStoreService.storeEmbeddings(storeEmbeddingBo);
+                for (KnowledgeFragment old : oldFragments) {
+                    if (StringUtils.isNotBlank(old.getFid())) {
+                        vectorStoreService.removeByFid(old.getFid(), String.valueOf(knowledgeId));
+                    }
+                }
+            } catch (Exception vectorError) {
+                for (String newFid : fids) {
+                    try {
+                        vectorStoreService.removeByFid(newFid, String.valueOf(knowledgeId));
+                    } catch (Exception cleanupError) {
+                        log.error("补偿删除新向量失败, kid={}, fid={}", knowledgeId, newFid, cleanupError);
+                    }
+                }
+                throw vectorError;
+            }
+
+            knowledgeFragmentMapper.delete(Wrappers.<KnowledgeFragment>lambdaQuery().eq(KnowledgeFragment::getDocId, docId));
+            knowledgeFragmentMapper.insertBatch(knowledgeFragmentList);
+            knowledgeRetrievalService.invalidateKnowledge(String.valueOf(knowledgeId));

            attach.setStatus(KnowledgeAttachStatus.COMPLETED.getCode()); // 已完成
            baseMapper.updateById(attach);
@@ -223,4 +293,22 @@ public class KnowledgeAttachServiceImpl implements IKnowledgeAttachService {
            baseMapper.updateById(attach);
        }
    }
+
+    @Override
+    public KnowledgeReparseVo reparseKnowledge(Long knowledgeId) {
+        List<KnowledgeAttach> attachments = baseMapper.selectList(
+            Wrappers.<KnowledgeAttach>lambdaQuery().eq(KnowledgeAttach::getKnowledgeId, knowledgeId));
+        int submitted = 0;
+        int skipped = 0;
+        IKnowledgeAttachService proxy = SpringUtils.getBean(IKnowledgeAttachService.class);
+        for (KnowledgeAttach attachment : attachments) {
+            if (KnowledgeAttachStatus.PARSING.getCode().equals(attachment.getStatus())) {
+                skipped++;
+            } else {
+                proxy.parse(attachment.getId());
+                submitted++;
+            }
+        }
+        return new KnowledgeReparseVo(submitted, skipped, attachments.size());
+    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/KnowledgeFragmentServiceImpl.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/KnowledgeFragmentServiceImpl.java
@@ -40,6 +40,7 @@ public class KnowledgeFragmentServiceImpl implements IKnowledgeFragmentService {
    private final IKnowledgeInfoService knowledgeInfoService;
    private final IChatModelService chatModelService;
    private final KnowledgeRetrievalService knowledgeRetrievalService;
+    private final org.ruoyi.service.vector.VectorStoreService vectorStoreService;

    /**
     * 查询知识片段
@@ -114,7 +115,11 @@ public class KnowledgeFragmentServiceImpl implements IKnowledgeFragmentService {
    public Boolean updateByBo(KnowledgeFragmentBo bo) {
        KnowledgeFragment update = MapstructUtils.convert(bo, KnowledgeFragment.class);
        validEntityBeforeSave(update);
-        return baseMapper.updateById(update) > 0;
+        boolean updated = baseMapper.updateById(update) > 0;
+        if (updated && update.getKnowledgeId() != null) {
+            knowledgeRetrievalService.invalidateKnowledge(String.valueOf(update.getKnowledgeId()));
+        }
+        return updated;
    }

    /**
@@ -136,6 +141,14 @@ public class KnowledgeFragmentServiceImpl implements IKnowledgeFragmentService {
        if(isValid){
            //TODO 做一些业务上的校验,判断是否需要校验
        }
+        // 删除 DB 片段前，同步删除向量库中对应向量
+        List<KnowledgeFragment> fragments = baseMapper.selectByIds(ids);
+        for (KnowledgeFragment fragment : fragments) {
+            if (StringUtils.isNotBlank(fragment.getFid()) && fragment.getKnowledgeId() != null) {
+                vectorStoreService.removeByFid(fragment.getFid(), String.valueOf(fragment.getKnowledgeId()));
+                knowledgeRetrievalService.invalidateKnowledge(String.valueOf(fragment.getKnowledgeId()));
+            }
+        }
        return baseMapper.deleteByIds(ids) > 0;
    }

--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/KnowledgeInfoServiceImpl.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/KnowledgeInfoServiceImpl.java
@@ -16,6 +16,10 @@ import org.ruoyi.mapper.knowledge.KnowledgeAttachMapper;
 import org.ruoyi.mapper.knowledge.KnowledgeInfoMapper;
 import org.ruoyi.service.knowledge.IKnowledgeInfoService;
 import org.springframework.stereotype.Service;
+import org.springframework.transaction.annotation.Transactional;
+import org.ruoyi.service.retrieval.KnowledgeRetrievalService;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;
+import org.ruoyi.common.core.service.OssService;

 import java.util.List;
 import java.util.Map;
@@ -36,6 +40,12 @@ public class KnowledgeInfoServiceImpl implements IKnowledgeInfoService {

    private final KnowledgeAttachMapper knowledgeAttachMapper;

+    private final org.ruoyi.mapper.knowledge.KnowledgeFragmentMapper knowledgeFragmentMapper;
+
+    private final org.ruoyi.service.vector.VectorStoreService vectorStoreService;
+    private final KnowledgeRetrievalService knowledgeRetrievalService;
+    private final OssService ossService;
+
    /**
     * 查询知识库
     *
@@ -97,10 +107,14 @@ public class KnowledgeInfoServiceImpl implements IKnowledgeInfoService {
     */
    private void fillDocumentCount(List<KnowledgeInfoVo> records) {
        if (records == null || records.isEmpty()) return;
-        for (KnowledgeInfoVo vo : records) {
-            int count = knowledgeAttachMapper.countByKnowledgeId(vo.getId());
-            vo.setDocumentCount(count);
+        List<Long> ids = records.stream().map(KnowledgeInfoVo::getId).toList();
+        Map<Long, Integer> counts = new java.util.HashMap<>();
+        for (Map<String, Object> row : knowledgeAttachMapper.countByKnowledgeIds(ids)) {
+            Number kid = (Number) (row.get("knowledgeId") != null ? row.get("knowledgeId") : row.get("knowledgeid"));
+            Number count = (Number) (row.get("documentCount") != null ? row.get("documentCount") : row.get("documentcount"));
+            if (kid != null && count != null) counts.put(kid.longValue(), count.intValue());
        }
+        records.forEach(vo -> vo.setDocumentCount(counts.getOrDefault(vo.getId(), 0)));
    }

    /**
@@ -130,13 +144,20 @@ public class KnowledgeInfoServiceImpl implements IKnowledgeInfoService {
    public Boolean updateByBo(KnowledgeInfoBo bo) {
        KnowledgeInfo update = MapstructUtils.convert(bo, KnowledgeInfo.class);
        validEntityBeforeSave(update);
-        return baseMapper.updateById(update) > 0;
+        boolean updated = baseMapper.updateById(update) > 0;
+        if (updated) knowledgeRetrievalService.invalidateKnowledge(String.valueOf(bo.getId()));
+        return updated;
    }

    /**
     * 保存前的数据校验
     */
    private void validEntityBeforeSave(KnowledgeInfo entity){
+        int blockSize = entity.getTextBlockSize() == null
+            ? DocumentSplitConfig.DEFAULT_BLOCK_SIZE : entity.getTextBlockSize().intValue();
+        int overlap = entity.getOverlapChar() == null
+            ? DocumentSplitConfig.DEFAULT_OVERLAP : entity.getOverlapChar().intValue();
+        new DocumentSplitConfig(entity.getSeparator(), blockSize, overlap, "");
        //TODO 做一些数据校验,如唯一约束
    }

@@ -148,10 +169,33 @@ public class KnowledgeInfoServiceImpl implements IKnowledgeInfoService {
     * @return 是否删除成功
     */
    @Override
+    @Transactional(rollbackFor = Exception.class)
    public Boolean deleteWithValidByIds(Collection<Long> ids, Boolean isValid) {
        if(isValid){
            //TODO 做一些业务上的校验,判断是否需要校验
        }
+        for (Long kid : ids) {
+            KnowledgeInfo info = baseMapper.selectById(kid);
+            // 1. 删除向量库中该知识库的所有向量（按文档逐个清理，三种向量库行为一致）
+            List<org.ruoyi.domain.entity.knowledge.KnowledgeAttach> attaches = knowledgeAttachMapper.selectList(
+                Wrappers.lambdaQuery(org.ruoyi.domain.entity.knowledge.KnowledgeAttach.class)
+                    .eq(org.ruoyi.domain.entity.knowledge.KnowledgeAttach::getKnowledgeId, kid));
+            vectorStoreService.removeById(String.valueOf(kid), info == null ? null : info.getVectorModel());
+            List<Long> ossIds = attaches.stream()
+                    .map(org.ruoyi.domain.entity.knowledge.KnowledgeAttach::getOssId)
+                    .filter(java.util.Objects::nonNull).toList();
+            if (!ossIds.isEmpty()) {
+                for (Long ossId : ossIds) {
+                    ossService.deleteFile(ossId);
+                }
+            }
+            // 2. 删除该知识库下的附件与片段记录
+            knowledgeAttachMapper.delete(Wrappers.lambdaQuery(org.ruoyi.domain.entity.knowledge.KnowledgeAttach.class)
+                .eq(org.ruoyi.domain.entity.knowledge.KnowledgeAttach::getKnowledgeId, kid));
+            knowledgeFragmentMapper.delete(Wrappers.lambdaQuery(org.ruoyi.domain.entity.knowledge.KnowledgeFragment.class)
+                .eq(org.ruoyi.domain.entity.knowledge.KnowledgeFragment::getKnowledgeId, kid));
+            knowledgeRetrievalService.invalidateKnowledge(String.valueOf(kid));
+        }
        return baseMapper.deleteByIds(ids) > 0;
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/CodeFileLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/CodeFileLoader.java
@@ -3,13 +3,16 @@ package org.ruoyi.service.knowledge.impl.loader;
 import lombok.AllArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;
 import org.ruoyi.service.knowledge.TextSplitter;
+import org.ruoyi.common.core.exception.ServiceException;
 import org.springframework.stereotype.Component;

 import java.io.BufferedReader;
 import java.io.IOException;
 import java.io.InputStream;
 import java.io.InputStreamReader;
+import java.nio.charset.StandardCharsets;
 import java.util.List;

@Component
@@ -21,20 +24,20 @@ public class CodeFileLoader implements ResourceLoader {
    @Override
    public String getContent(InputStream inputStream) {
        StringBuffer stringBuffer = new StringBuffer();
-        try (InputStreamReader reader = new InputStreamReader(inputStream);
+        try (InputStreamReader reader = new InputStreamReader(inputStream, StandardCharsets.UTF_8);
             BufferedReader bufferedReader = new BufferedReader(reader)) {
            String line;
            while ((line = bufferedReader.readLine()) != null) {
                stringBuffer.append(line).append("\n");
            }
        } catch (IOException e) {
-            e.printStackTrace();
+            throw new ServiceException("读取代码文件失败", e);
        }
        return stringBuffer.toString();
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
-        return textSplitter.split(content, kid);
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
+        return textSplitter.split(content, config);
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/CsvFileLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/CsvFileLoader.java
@@ -1,6 +1,7 @@
 package org.ruoyi.service.knowledge.impl.loader;

 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;

 import java.io.InputStream;
 import java.util.List;
@@ -12,7 +13,7 @@ public class CsvFileLoader implements ResourceLoader {
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
        return null;
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/ExcelFileLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/ExcelFileLoader.java
@@ -6,6 +6,7 @@ import dev.langchain4j.data.document.parser.apache.tika.ApacheTikaDocumentParser
 import lombok.AllArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;
 import org.ruoyi.service.knowledge.TextSplitter;
 import org.springframework.stereotype.Component;

@@ -38,7 +39,7 @@ public class ExcelFileLoader implements ResourceLoader {
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
-        return textSplitter.split(content, kid);
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
+        return textSplitter.split(content, config);
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/FolderLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/FolderLoader.java
@@ -1,6 +1,7 @@
 package org.ruoyi.service.knowledge.impl.loader;

 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;

 import java.io.InputStream;
 import java.util.List;
@@ -12,7 +13,7 @@ public class FolderLoader implements ResourceLoader {
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
        return null;
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/GithubLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/GithubLoader.java
@@ -1,6 +1,7 @@
 package org.ruoyi.service.knowledge.impl.loader;

 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;

 import java.io.InputStream;
 import java.util.List;
@@ -12,7 +13,7 @@ public class GithubLoader implements ResourceLoader {
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
        return null;
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/JsonFileLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/JsonFileLoader.java
@@ -1,6 +1,7 @@
 package org.ruoyi.service.knowledge.impl.loader;

 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;

 import java.io.InputStream;
 import java.util.List;
@@ -12,7 +13,7 @@ public class JsonFileLoader implements ResourceLoader {
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
        return null;
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/MarkDownFileLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/MarkDownFileLoader.java
@@ -3,13 +3,16 @@ package org.ruoyi.service.knowledge.impl.loader;
 import lombok.AllArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;
 import org.ruoyi.service.knowledge.TextSplitter;
+import org.ruoyi.common.core.exception.ServiceException;
 import org.springframework.stereotype.Component;

 import java.io.BufferedReader;
 import java.io.IOException;
 import java.io.InputStream;
 import java.io.InputStreamReader;
+import java.nio.charset.StandardCharsets;
 import java.util.List;

@Component
@@ -21,20 +24,20 @@ public class MarkDownFileLoader implements ResourceLoader {
    @Override
    public String getContent(InputStream inputStream) {
        StringBuffer stringBuffer = new StringBuffer();
-        try (InputStreamReader reader = new InputStreamReader(inputStream);
+        try (InputStreamReader reader = new InputStreamReader(inputStream, StandardCharsets.UTF_8);
             BufferedReader bufferedReader = new BufferedReader(reader)) {
            String line;
            while ((line = bufferedReader.readLine()) != null) {
                stringBuffer.append(line).append("\n");
            }
        } catch (IOException e) {
-            e.printStackTrace();
+            throw new ServiceException("读取 Markdown 文件失败", e);
        }
        return stringBuffer.toString();
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
-        return textSplitter.split(content, kid);
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
+        return textSplitter.split(content, config);
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/PdfFileLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/PdfFileLoader.java
@@ -6,6 +6,7 @@ import org.apache.pdfbox.io.RandomAccessReadBuffer;
 import org.apache.pdfbox.pdmodel.PDDocument;
 import org.apache.pdfbox.text.PDFTextStripper;
 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;
 import org.ruoyi.service.knowledge.TextSplitter;
 import org.springframework.stereotype.Component;

@@ -32,7 +33,7 @@ public class PdfFileLoader implements ResourceLoader {
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
-        return characterTextSplitter.split(content, kid);
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
+        return characterTextSplitter.split(content, config);
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/TextFileLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/TextFileLoader.java
@@ -3,7 +3,9 @@ package org.ruoyi.service.knowledge.impl.loader;
 import lombok.AllArgsConstructor;
 import lombok.extern.slf4j.Slf4j;
 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;
 import org.ruoyi.service.knowledge.TextSplitter;
+import org.ruoyi.common.core.exception.ServiceException;
 import org.springframework.stereotype.Component;

 import java.io.BufferedReader;
@@ -22,18 +24,16 @@ public class TextFileLoader implements ResourceLoader {

    @Override
    public String getContent(InputStream inputStream) {
-        String stringBuffer = "";
        try (InputStreamReader reader = new InputStreamReader(inputStream, StandardCharsets.UTF_8);
             BufferedReader bufferedReader = new BufferedReader(reader)) {
-            stringBuffer = bufferedReader.lines().collect(Collectors.joining());
+            return bufferedReader.lines().collect(Collectors.joining("\n"));
        } catch (IOException e) {
-            e.printStackTrace();
+            throw new ServiceException("读取文本文件失败", e);
        }
-        return stringBuffer;
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
-        return textSplitter.split(content, kid);
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
+        return textSplitter.split(content, config);
    }
 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/WordLoader.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/loader/WordLoader.java
@@ -5,6 +5,7 @@ import lombok.extern.slf4j.Slf4j;
 import org.apache.poi.xwpf.extractor.XWPFWordExtractor;
 import org.apache.poi.xwpf.usermodel.XWPFDocument;
 import org.ruoyi.service.knowledge.ResourceLoader;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;
 import org.ruoyi.service.knowledge.TextSplitter;
 import org.springframework.stereotype.Component;

@@ -32,8 +33,8 @@ public class WordLoader implements ResourceLoader {
    }

    @Override
-    public List<String> getChunkList(String content, String kid) {
-        return textSplitter.split(content, kid);
+    public List<String> getChunkList(String content, DocumentSplitConfig config) {
+        return textSplitter.split(content, config);
    }

 }
--- a/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/split/CharacterTextSplitter.java
+++ b/ruoyi-modules/ruoyi-chat/src/main/java/org/ruoyi/service/knowledge/impl/split/CharacterTextSplitter.java
@@ -1,82 +1,17 @@
 package org.ruoyi.service.knowledge.impl.split;

-import lombok.AllArgsConstructor;
-import lombok.extern.slf4j.Slf4j;
-import org.ruoyi.common.core.utils.StringUtils;
-import org.ruoyi.domain.vo.knowledge.KnowledgeInfoVo;
-import org.ruoyi.service.knowledge.IKnowledgeInfoService;
+import org.ruoyi.service.knowledge.DocumentSplitConfig;
 import org.ruoyi.service.knowledge.TextSplitter;
 import org.springframework.context.annotation.Primary;
 import org.springframework.stereotype.Component;

-import java.util.ArrayList;
-import java.util.Arrays;
 import java.util.List;

@Component
-@Slf4j
@Primary
-@AllArgsConstructor
 public class CharacterTextSplitter implements TextSplitter {
-
-    private final IKnowledgeInfoService knowledgeInfoService;
-
    @Override
-    public List<String> split(String content, String kid) {
-        // 默认配置值
-        String knowledgeSeparator = "#";
-        int textBlockSize = 1000;
-        int overlapChar = 50;
-
-        // 根据知识库ID查询配置，覆盖默认值
-        if (StringUtils.isNotBlank(kid)) {
-            try {
-                KnowledgeInfoVo info = knowledgeInfoService.queryById(Long.parseLong(kid));
-                if (info != null) {
-                    if (StringUtils.isNotBlank(info.getSeparator())) {
-                        knowledgeSeparator = info.getSeparator();
-                    }
-                    if (info.getTextBlockSize() != null && info.getTextBlockSize() > 0) {
-                        textBlockSize = info.getTextBlockSize().intValue();
-                    }
-                    if (info.getOverlapChar() != null && info.getOverlapChar() > 0) {
-                        overlapChar = info.getOverlapChar().intValue();
-                    }
-                }
-            } catch (Exception e) {
-                log.warn("查询知识库配置失败，使用默认配置, kid={}", kid, e);
-            }
-        }
-
-        List<String> chunkList = new ArrayList<>();
-        if (content.contains(knowledgeSeparator) && StringUtils.isNotBlank(knowledgeSeparator)) {
-            // 按自定义分隔符切分
-            String[] chunks = content.split(knowledgeSeparator);
-            chunkList.addAll(Arrays.asList(chunks));
-        } else {
-            int indexMin = 0;
-            int len = content.length();
-            int i = 0;
-            int right = 0;
-            while (true) {
-                if (len > right) {
-                    int begin = i * textBlockSize - overlapChar;
-                    if (begin < indexMin) {
-                        begin = indexMin;
-                    }
-                    int end = textBlockSize * (i + 1) + overlapChar;
-                    if (end > len) {
-                        end = len;
-                    }
-                    String chunk = content.substring(begin, end);
-                    chunkList.add(chunk);
-                    i++;
-                    right = right + textBlockSize;
-                } else {
-                    break;
-                }
-            }
-        }
-        return chunkList;
+    public List<String> split(String content, DocumentSplitConfig config) {
+        return SplitterSupport.split(content, config, SplitterSupport::paragraphs);
    }
 }
--- a/Show More
+++ b/Show More