【Spring AI】基于SpringAI+Vue3+ElementPlus的QA系统实现一

• Indexing pipeline 阶段主要是将结构化或者非结构化的数据或文档进行加载和解析、chunk切分、文本向量化并保存到向量数据库。

• RAG 阶段主要包括将 promp t文本内容转为向量、从向量数据库检索内容、对检索后的文档 chunk 进行重排和 prompt 重写、最后调用大模型进行结果的生成。

• 已有自己的向量库，如 Redis、Faiss。想了解更多参考我的博客：

  【LLM】Redis 作为向量库入门指南_redis 向量数据库-CSDN博客文章浏览阅读1k次，点赞25次，收藏12次。这篇文章将介绍基于 RedisSearch 的Redis向量库实现。通过阅读本篇文章，你将学习到如何创建向量索引，如何存储和更新向量，如何进行向量搜索，如何使用阿里百炼 Embedding Model 文本向量化，如何集成到 spring boot 中并实现向量的存储和搜索等。_redis 向量数据库https://blog.csdn.net/u013176571/article/details/145122380

• 已安装 18.3 或更高版本的 Node.js ，未安装进入 官网下载 ，LTS 为长期支持版，Current 为最新功能版。

• 使用过Spring AI，可参考我的这篇博客：

  【Spring AI】Spring AI Alibaba的简单使用_spring-ai-alibaba-starter-CSDN博客文章浏览阅读696次，点赞7次，收藏6次。项目中引入Spring AI、Spring AI Alibaba踩坑笔记，并实现简单的几种聊天模式。_spring-ai-alibaba-starterhttps://blog.csdn.net/u013176571/article/details/144488475

<dependency><groupId>org.springframework.ai</groupId><artifactId>spring-ai-core</artifactId><version>${spring-ai-core.version}</version>
</dependency><dependency><groupId>org.springframework.ai</groupId><artifactId>spring-ai-openai</artifactId><version>${spring-ai-core.version}</version>
</dependency>

<repositories><repository><id>maven2</id><name>maven2</name><url>https://repo1.maven.org/maven2/</url><snapshots><enabled>false</enabled></snapshots></repository><repository><id>spring-milestones</id><name>Spring Milestones</name><url>https://repo.spring.io/milestone</url><snapshots><enabled>false</enabled></snapshots></repository>
</repositories>

spring:ai:openai:api-key: your-api-keybase-url: https://api.deepseek.comchat:model: deepseek-chatchat:client:enabled: false

@Bean
public EmbeddingModel embeddingModel() {return new OpenAiEmbeddingModel(// embeddingBaseUrl 为阿里百炼的接口 url ，值为：https://dashscope.aliyuncs.com/compatible-mode// embeddingApiKey 为阿里百炼上申请的 api-key 。new OpenAiApi(embeddingBaseUrl, embeddingApiKey),MetadataMode.EMBED,OpenAiEmbeddingOptions.builder().model("text-embedding-v3")  // 选用的嵌入模型.dimensions(1024)            // 向量维度.build(),RetryUtils.DEFAULT_RETRY_TEMPLATE);
}

@Bean
public ChatModel chatModel() {// baseUrl 为 DeepSeek 接口 url，值为：https://api.deepseek.com// apiKey 为申请的秘钥，defaultChatModel 为聊天模型。return new OpenAiChatModel(new OpenAiApi(baseUrl, apiKey), OpenAiChatOptions.builder().model("deepseek-chat")    // 选用的聊天模型.temperature(0.7d)         // .build());
}

@Bean
public VectorStore vectorStore() {return RedisVectorStore.builder(jedisPooled, embeddingModel).indexName(indexName).prefix(prefix).embeddingFieldName("embedding")      // 向量字段名，默认 embedding.contentFieldName("content")          // 存储的原始文本内容.initializeSchema(initializeSchema)   // 是否初始化索引配置信息.batchingStrategy(batchingStrategy).metadataFields(RedisVectorStore.MetadataField.text(SpringAIConst.VectorStore.METADATA_DOC_ID),   // 存储的元数据信息RedisVectorStore.MetadataField.text(SpringAIConst.VectorStore.METADATA_DOC_NAME)  // 存储的元数据信息).build();
}

public class RecursiveCharacterTextSplitter implements TextSplitterIntf {/*** 分隔符列表，用于拆分文本* 中文："\n\n", "。", "！", "？", "，", "；"* 英文："\n\n", "\n", " "*/private List<String> separators = Arrays.asList("\n\n", "。", "！", "？", "，", "；");/*** 每个块的最大大小*/private int chunkSize = 250;/*** 块之间的重叠大小，用来维持上下文关系*/private int chunkOverlap = 30;public RecursiveCharacterTextSplitter() {}public RecursiveCharacterTextSplitter(int chunkSize, int chunkOverlap) {this.chunkSize = chunkSize;this.chunkOverlap = chunkOverlap;if (chunkOverlap >= chunkSize) {throw new IllegalArgumentException("chunkOverlap must be smaller than chunkSize");}}public RecursiveCharacterTextSplitter(List<String> separators, int chunkSize, int chunkOverlap) {this.separators = separators;this.chunkSize = chunkSize;this.chunkOverlap = chunkOverlap;if (chunkOverlap >= chunkSize) {throw new IllegalArgumentException("chunkOverlap must be smaller than chunkSize");}}/*** Spring ai Document 拆分** @param documents* @param metaData  文档元数据信息* @return*/@Overridepublic List<Document> split(List<Document> documents, List<Map<String, Object>> metaData) {List<JSONObject> docList = this.beforeSplit(documents, metaData);List<Document> chunkedDocuments = new ArrayList<>();docList.forEach(item -> {Document document = item.getObject("document", Document.class);Map<String, Object> singleMetaData = item.getObject("metaData", Map.class);List<String> chunks = new ArrayList<>();splitRecursive(document.getContent(), chunks);chunks.forEach(chunk -> {chunkedDocuments.add(new Document(chunk, singleMetaData));});});return chunkedDocuments;}/*** 递归地拆分文本** @param text* @param chunks*/private void splitRecursive(String text, List<String> chunks) {if (text.length() <= chunkSize) {chunks.add(text);return;}String separator = findBestSeparator(text);if (separator == null) {chunks.add(text);return;}String[] parts = text.split(separator, -1);StringBuilder currentChunk = new StringBuilder();int currentLength = 0;for (String part : parts) {if (currentLength + part.length() > chunkSize) {if (currentChunk.length() > 0) {chunks.add(currentChunk.toString());// 通过回退chunkOverlap的距离创建来重叠部分int overlapStart = Math.max(0, currentChunk.length() - chunkOverlap);currentChunk = new StringBuilder(currentChunk.substring(overlapStart));currentLength = currentChunk.length();}}currentChunk.append(part).append(separator);currentLength += part.length() + separator.length();}if (currentChunk.length() > 0) {chunks.add(currentChunk.toString());}}/*** 查找文本中第一个出现的分隔符** @param text* @return*/private String findBestSeparator(String text) {for (String separator : separators) {if (text.contains(separator)) {return separator;}}return null;}
}

@Bean
public BatchingStrategy tokenCountBatchingStrategy() {return new TokenCountBatchingStrategy(EncodingType.CL100K_BASE,  // 指定用于分词的编码类型8000,                      // 最大 Token 数量0.1                        // 设置预留百分比（默认 10%）);
}

    /*** 文档入库** @param inputStream 文档输入流* @param docHash     文档Hash* @param docName     文档名* @throws Exception*/public void addDocument(InputStream inputStream, String docHash, String docName) throws Exception {// 校验文档是否已经存在if (redisService.hasKey(CommonConst.RAG_KEY_PREFIX + docHash)) {return;}int docSize = inputStream.available();// 文件流转为 DocumentTextReader textReader = new TextReader(new InputStreamResource(inputStream));List<Document> documents = textReader.get();// 文档拆分TextSplitterIntf textSplitter = new RecursiveCharacterTextSplitter(250, 30);List<Document> newDocuments = textSplitter.split(documents, List.of(Map.of(SpringAIConst.VectorStore.METADATA_DOC_ID, docHash,SpringAIConst.VectorStore.METADATA_DOC_NAME, docName)));// 向量入库vectorStore.add(newDocuments);// 将文档信息存入 RedisJSONObject docJson = new JSONObject();docJson.put(SpringAIConst.VectorStore.METADATA_DOC_NAME, docName);docJson.put("docSize", docSize);docJson.put("docChunks", newDocuments.size());redisService.setObject(CommonConst.RAG_KEY_PREFIX + docHash, docJson);}

    /*** 计算文件 Hash 值** @param inputStream 输入流* @param algorithm   算法，可选，MD5、SHA-1、SHA-256* @return* @throws Exception*/public static String calculateFileHash(InputStream inputStream, String algorithm) throws Exception {// 创建MessageDigest实例algorithm = StringUtils.isEmpty(algorithm) ? CommonConst.SHA_256 : algorithm;MessageDigest digest = MessageDigest.getInstance(algorithm);// 打开文件输入流try (inputStream) {byte[] buffer = new byte[1024];int numRead;// 读取文件内容并更新摘要while ((numRead = inputStream.read(buffer)) != -1) {digest.update(buffer, 0, numRead);}// 完成哈希计算byte[] hashBytes = digest.digest();// 将哈希值转换为十六进制字符串StringBuilder sb = new StringBuilder();for (byte b : hashBytes) {sb.append(String.format("%02x", b));}return sb.toString();}}

    /*** 带聊天记忆流式返回** @param accessKey 聊天访问key，每次打开聊天页面重置* @param inputInfo 聊天输入内容* @param docIds    参考文档ID，多个用英文','号隔开* @return*/@GetMapping("/memoryStreamWithApi")public Flux<ServerSentEvent<String>> memoryStreamWithApi(String accessKey, String inputInfo, String docIds) {// 接收并校验参数CommonUtil.checkAndThrow(accessKey, "您没有该功能访问权限！");CommonUtil.checkAndThrow(inputInfo, "请输入内容！");// 调用模型查询ChatClient.ChatClientRequestSpec chatClientRequestSpec = ChatClient.builder(chatModel).defaultAdvisors(new MessageChatMemoryAdvisor(new InMemoryChatMemory())).build().prompt().user(inputInfo).advisors(advisor -> advisor.param(CHAT_MEMORY_CONVERSATION_ID_KEY, accessKey).param(CHAT_MEMORY_RETRIEVE_SIZE_KEY, 5));// 先从向量库查询近似数据if (StringUtils.isNotEmpty(docIds)) {// 组装过滤条件，根据文档ID过滤，等价于 SQL 的 WHERE docId IN (docId1,docId2...)List<String> docIdList = Arrays.asList(docIds.split(","));Filter.Expression expression = new Filter.Expression(Filter.ExpressionType.IN,new Filter.Key(SpringAIConst.VectorStore.METADATA_DOC_ID), new Filter.Value(docIdList));Advisor advisor = new QuestionAnswerAdvisor(vectorStore,SearchRequest.builder().query(inputInfo).filterExpression(expression).topK(6).similarityThreshold(0.8d).build());chatClientRequestSpec.advisors(advisor);}return chatClientRequestSpec.stream().chatResponse().map(response -> ServerSentEvent.<String>builder().data(response.getResult().getOutput().getContent()).build());}

• Spring AI RAG 文档
• Spring AI Alibaba RAG 文档
• Vue3 文档
• Element Plus 文档
• markdown-it 文档

9.1.1 阿里通义

Token 是模型用来表示自然语言文本的基本单位，可以直观地理解为“字”或“词”。

不同的大模型切分Token 的方法可能不同。可以本地运行 tokenizer 来估计文本的 Token 量。

9.1.2 DeepSeek

一般情况下模型中 token 和字数的换算比例大致如下：