企业级RAG知识库系统：从文档到智能问答的全链路实现

graph TB
    subgraph 用户端
        A[Vue3 + Naive UI]
    end

    subgraph API层
        B[Spring Boot REST API]
        C[WebSocket Handler]
    end

    subgraph 异步流水线
        D[Kafka Producer]
        E[Tika 文档解析]
        F[豆包 Embedding 向量化]
        G[Elasticsearch 索引]
    end

    subgraph RAG检索
        H[用户提问]
        I[Embedding转换]
        J[KNN向量召回]
        K[BM25关键词匹配]
        L[分数融合重排序]
        M[增强Prompt构建]
        N[DeepSeek生成回答]
    end

    subgraph 基础设施
        O[(MySQL)]
        P[(Redis)]
        Q[(MinIO)]
        R[(Elasticsearch)]
    end

    A --> B
    A --> C
    B --> D
    D --> E --> F --> G --> R
    H --> I --> J
    H --> K
    J --> L --> M --> N
    B --> O
    B --> P
    B --> Q

@Service
public class ChunkUploadService {

    private static final String UPLOAD_KEY = "upload:";

    /**
     * 标记分片已上传
     * 1000个分片仅需 125 字节
     */
    public void markUploaded(String fileId, int chunkIndex) {
        String key = UPLOAD_KEY + fileId;
        redisTemplate.opsForValue().setBit(key, chunkIndex, true);
    }

    /**
     * 检查分片状态（秒传判断）
     */
    public boolean isUploaded(String fileId, int chunkIndex) {
        return Boolean.TRUE.equals(
            redisTemplate.opsForValue()
                .getBit(UPLOAD_KEY + fileId, chunkIndex)
        );
    }

    /**
     * 统计已上传分片数
     */
    public long countUploaded(String fileId) {
        return redisTemplate.execute(
            (RedisCallback<Long>) con ->
                con.stringCommands()
                    .bitCount(
                        (UPLOAD_KEY + fileId).getBytes()
                    )
        );
    }
}

graph LR
    A[用户提问] --> B[豆包Embedding<br/>转为2048维向量]
    A --> C[IK分词器<br/>提取关键词]
    B --> D[Elasticsearch KNN<br/>向量语义召回]
    C --> E[ES Match Query<br/>关键词匹配]
    D --> F[分数融合 + 重排序]
    E --> F
    F --> G[Top-K 相关片段]

SearchRequest request = SearchRequest.of(sr -> sr
    .index("knowledge_base")
    .size(10)
    .query(q -> q.bool(b -> b
        // 向量语义搜索
        .should(s -> s.knn(k -> k
            .field("embedding")
            .queryVector(embeddingVector)
            .numCandidates(50)
            .k(10)
        ))
        // BM25 关键词搜索
        .should(s -> s.multiMatch(m -> m
            .fields("title", "content")
            .query(queryText)
            .type(TextQueryType.BestFields)
        ))
        .minimumShouldMatch("1")
    ))
);

sequenceDiagram
    participant 用户
    participant API
    participant Kafka
    participant Consumer
    participant ES

    用户->>API: 上传文档
    API->>Kafka: 发送处理事件
    API-->>用户: 200 OK（响应 < 200ms）
    Note over Kafka,ES: 异步流水线
    Kafka->>Consumer: 消费事件
    Consumer->>Consumer: Tika 解析文档
    Consumer->>Consumer: 豆包 Embedding 向量化
    Consumer->>ES: 写入索引

// Producer：上传即返回
@PostMapping("/upload")
public Result<String> upload(MultipartFile file) {
    String fileId = generateId();
    minioService.upload(file, fileId);
    kafkaTemplate.send("doc-process",
        new DocProcessEvent(fileId, file.getName()));
    return Result.ok(fileId); // < 200ms
}

// Consumer：异步处理
@KafkaListener(topics = "doc-process")
public void process(DocProcessEvent event) {
    File file = minioService.download(event.getFileId());
    String text = tikaParser.parse(file);
    float[] vector = embeddingClient.embed(text);
    esService.index(event.getFileId(), text, vector);
}

@Service
public class ChatHistoryService {

    private static final Duration TTL = Duration.ofDays(7);

    public void append(String sessionId, ChatMessage msg) {
        String key = "chat:" + sessionId;
        redisTemplate.opsForList().rightPush(key, msg);
        redisTemplate.expire(key, TTL);
    }

    public List<ChatMessage> getRecent(String sessionId, int rounds) {
        String key = "chat:" + sessionId;
        Long size = redisTemplate.opsForList().size(key);
        if (size == null || size == 0) return List.of();
        int end = size.intValue() - 1;
        int start = Math.max(0, end - rounds * 2 + 1);
        return redisTemplate.opsForList().range(key, start, end);
    }
}