化)
一、混合檢索向量 關鍵詞1. 純向量檢索的局限語義相近但答案靠后向量相似度高不代表文檔真正包含答案。精確關鍵詞匹配弱用戶搜索特定術語時向量檢索可能漏掉精確匹配的文檔。多路召回更可靠結合向量檢索和關鍵詞檢索取并集后融合排序。2. 混合檢索的實現(xiàn)以PGVector為例-- PGVector 全文檢索 向量檢索 SQL SELECT * FROM knowledge_base WHERE tsv to_tsquery(chinese, :query) -- 關鍵詞匹配 ORDER BY ts_rank(tsv, to_tsquery(chinese, :query)) DESC, -- 關鍵詞相關性 embedding :query_vector -- 向量相似度 LIMIT :topK;3. Java端混合檢索邏輯核心思路向量檢索與關鍵詞檢索并行結果融合去重RRF排序。Service public class HybridSearchService { Resource private VectorStore vectorStore; Resource private JdbcTemplate jdbcTemplate; public ListDocument hybridSearch(String query, int topK) { // 1. 向量檢索 ListDocument vectorResults vectorStore.similaritySearch( SearchRequest.builder().query(query).topK(topK).build() ); // 2. 關鍵詞檢索PGVector全文搜索 String sql SELECT id, content, ts_rank(tsv, to_tsquery(chinese, ?)) AS rank FROM knowledge_base WHERE tsv to_tsquery(chinese, ?) ORDER BY rank DESC LIMIT ? ; ListDocument keywordResults jdbcTemplate.query(sql, ...); // 3. RRF融合排序Reciprocal Rank fusion return rrfMerge(vectorResults, keywordResults, topK); } /** * RRFReciprocal Rank Fusion融合排序 * 對兩個檢索結果按排名的倒數(shù)加權融合得到最終排序 */ private ListDocument rrfMerge( ListDocument vectorResults, ListDocument keywordResults, int topK) { double k 60.0; // RRF常量 MapString, Double scores new HashMap(); MapString, Document docMap new HashMap(); // 向量檢索結果按排名加權 for (int i 0; i vectorResults.size(); i) { String id vectorResults.get(i).getId(); double score 1.0 / (k i 1); scores.merge(id, score, Double::sum); docMap.put(id, vectorResults.get(i)); } // 關鍵詞檢索結果按排名加權 for (int i 0; i keywordResults.size(); i) { String id keywordResults.get(i).getId(); double score 1.0 / (k i 1); scores.merge(id, score, Double::sum); docMap.put(id, keywordResults.get(i)); } // 按融合分數(shù)降序取topK return scores.entrySet().stream() .sorted(Map.Entry.String, DoublecomparingByValue().reversed()) .limit(topK) .map(e - docMap.get(e.getKey())) .toList(); } }4. 集成到RAG流程GetMapping(/ask-hybrid) public String askHybrid(RequestParam String question) { // Step 1: 混合檢索 ListDocument docs hybridSearchService.hybridSearch(question, 10); // Step 2: 生成答案 String context buildContext(docs); return chatClient.prompt() .user(參考資料\n context \n\n問題 question) .call().content(); }二、Rerank重排序Rerank 是對檢索結果按查詢相關性進行二次精排的過程。粗檢索階段多召回topK20Rerank階段精排取Top5。1. DashScope Rerankqwen3-rerank通過阿里云 DashScope API 調用 qwen3-rerank 模型實現(xiàn)基于API的文檔重排序。Service public class DashScopeDocumentReranker implements DocumentPostProcessor { private final RestTemplate restTemplate; private final String apiKey; private final String modelName; private final int topN; public DashScopeDocumentReranker(String apiKey, String modelName, int topN) { this.restTemplate new RestTemplate(); this.apiKey apiKey; this.modelName modelName; this.topN topN; } Override public ListDocument process(Query query, ListDocument documents) { if (documents null || documents.size() 1) { return documents; } ListString texts documents.stream().map(Document::getText).toList(); MapString, Object body Map.of( model, modelName, query, query.text(), documents, texts, top_n, Math.min(topN, documents.size()) ); HttpHeaders headers new HttpHeaders(); headers.setContentType(MediaType.APPLICATION_JSON); headers.set(Authorization, Bearer apiKey); try { ResponseEntityRerankResponse response restTemplate.postForEntity( https://dashscope.aliyuncs.com/api/v1/services/rerank, new HttpEntity(body, headers), RerankResponse.class ); return reorder(documents, response.getBody()); } catch (Exception e) { log.warn(DashScope rerank調用失敗: {}, e.getMessage()); return documents; } } /** * 按rerank返回的相關性順序重排文檔結果已按relevance_score降序排列 */ private ListDocument reorder(ListDocument documents, RerankResponse response) { if (response null || response.results null || response.results.isEmpty()) { return documents; } ListDocument reranked new ArrayList(); for (RerankResult result : response.results) { if (result.index ! null result.index 0 result.index documents.size()) { reranked.add(documents.get(result.index)); } } return reranked.isEmpty() ? documents : reranked; } // qwen3-rerank API 響應結構 static class RerankResponse { public ListRerankResult results; } static class RerankResult { JsonProperty(relevance_score) public Double relevanceScore; public Integer index; } }2. 硅基流動 RerankBAAI/bge-reranker-v2-m3調用硅基流動平臺的 bge-reranker-v2-m3 模型進行文檔重排序。Service public class RerankService { private static final String API_URL https://api.siliconflow.cn/v1/rerank; private final RestTemplate restTemplate; private final ObjectMapper objectMapper; Value(${spring.ai.siliconflow-api-key}) private String apiKey; public RerankService(RestTemplateBuilder restTemplateBuilder) { this.restTemplate restTemplateBuilder .connectTimeout(Duration.ofSeconds(10)) .readTimeout(Duration.ofSeconds(30)) .build(); this.objectMapper new ObjectMapper(); } public ListDocument rerank(String query, ListDocument documents, int topN) { if (documents null || documents.isEmpty()) { return List.of(); } if (documents.size() 1) { return documents; } ListString texts documents.stream().map(Document::getText).toList(); MapString, Object body Map.of( model, BAAI/bge-reranker-v2-m3, query, query, documents, texts, top_n, Math.min(topN, documents.size()) ); try { HttpHeaders headers new HttpHeaders(); headers.setContentType(MediaType.APPLICATION_JSON); headers.set(Authorization, Bearer apiKey); HttpEntityMapString, Object request new HttpEntity(body, headers); ResponseEntityMap response restTemplate.postForEntity( API_URL, request, Map.class ); SuppressWarnings(unchecked) ListMapString, Object results (ListMapString, Object) response.getBody().get(results); if (results null || results.isEmpty()) { return documents; } return results.stream().map(r - { int index (int) r.get(index); Document doc documents.get(index); doc.getMetadata().put(rerankScore, r.get(relevance_score)); return doc; }).toList(); } catch (Exception e) { log.warn(硅基流動Rerank調用失敗: {}, e.getMessage()); return documents.stream().limit(topN).toList(); } } }3. LLM文檔重排器自定義實現(xiàn)Spring AI 2.0.1 沒有內置 reranker可以用 LLM 對檢索結果按相關性重新排序。/** * 基于LLM的文檔重排器實現(xiàn)DocumentPostProcessor接口 */ public class LlmDocumentReranker implements DocumentPostProcessor { private static final Pattern INDEX_PATTERN Pattern.compile(\\d); private final ChatClient chatClient; private final int maxDocPreviewLength; public LlmDocumentReranker(ChatClient.Builder chatClientBuilder, int maxDocPreviewLength) { this.chatClient chatClientBuilder.build(); this.maxDocPreviewLength maxDocPreviewLength; } Override public ListDocument process(Query query, ListDocument documents) { if (documents null || documents.size() 1) { return documents; } StringBuilder docList new StringBuilder(); for (int i 0; i documents.size(); i) { String text documents.get(i).getText(); if (text null) { text ; } if (text.length() maxDocPreviewLength) { text text.substring(0, maxDocPreviewLength) ...; } docList.append([).append(i).append(] ).append(text).append(\n\n); } String prompt 你是一個文檔相關性排序助手。請根據(jù)文檔與查詢的相關程度重新排序。 查詢%s 文檔列表 %s 請只返回文檔編號的排序結果用英文逗號分隔。 例如2,0,3,1 .formatted(query.text(), docList); String result chatClient.prompt().user(prompt).call().content(); return reorder(documents, result); } /** * 解析LLM返回的編號序列按順序重排文檔。解析失敗時返回原順序。 */ private ListDocument reorder(ListDocument documents, String llmOutput) { if (llmOutput null || llmOutput.isBlank()) { return documents; } Matcher matcher INDEX_PATTERN.matcher(llmOutput); ListInteger indices new ArrayList(); while (matcher.find()) { int idx Integer.parseInt(matcher.group()); if (idx 0 idx documents.size() !indices.contains(idx)) { indices.add(idx); } } if (indices.size() ! documents.size()) { return documents; } ListDocument reranked new ArrayList(documents.size()); for (int idx : indices) { reranked.add(documents.get(idx)); } return reranked; } }4. Rerank集成方式方式一手動編排GetMapping(/ask-rerank) public String askWithRerank(RequestParam String question) { // Step 1: 粗檢索多召回 ListDocument candidates vectorStore.similaritySearch( SearchRequest.builder().query(question).topK(20).build() ); // Step 2: Rerank精排取Top5 ListDocument reranked rerankService.rerank(question, candidates, 5); // Step 3: 生成答案 String context buildContext(reranked); return chatClient.prompt() .user(參考資料\n context \n\n問題 question) .call().content(); }方式二Advisor集成RetrievalAugmentationAdvisor ragAdvisor RetrievalAugmentationAdvisor.builder() .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .similarityThreshold(0.7) .topK(6) .build()) .documentPostProcessors(new DashScopeDocumentReranker(apiKey, qwen3-rerank, 3)) .build();三、Query改寫與多查詢擴展1. Query改寫的必要性用戶提問可能口語化、模糊或含指代詞直接用于檢索效果差。通過LLM改寫為更適合知識庫檢索的形式提升召回率。2. RewriteQueryTransformerSpring AI內置Spring AI 2.0 提供了內置的查詢改寫器可直接集成到 Advisor 中。Bean public RetrievalAugmentationAdvisor ragAdvisorWithRewrite( ChatClient.Builder chatClientBuilder, VectorStore vectorStore) { RewriteQueryTransformer rewriteTransformer RewriteQueryTransformer.builder() .chatClientBuilder(chatClientBuilder) .targetSearchSystem(vector store) .build(); return RetrievalAugmentationAdvisor.builder() .queryTransformers(rewriteTransformer) .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .topK(5) .build()) .build(); }3. 自定義多查詢改寫服務生成多個改寫版本分別檢索后取并集進一步提升召回率。Service public class QueryRewriteService { Resource private ChatClient ragManualChatClient; private final ObjectMapper objectMapper; public QueryRewriteService() { this.objectMapper new ObjectMapper(); } public ListString rewrite(String originalQuery, String conversationHistory) { String prompt String.format( 你是一個搜索查詢優(yōu)化專家。請將用戶的問題改寫為更適合在知識庫中檢索的形式。 要求 1. 生成3個不同角度的改寫版本 2. 用JSON數(shù)組格式返回 3. 保留原始問題的核心意圖 4. 如果有對話歷史結合上下文補全指代 對話歷史%s 用戶問題%s 請直接返回JSON數(shù)組不要包含其他文字。 , conversationHistory, originalQuery); String result ragManualChatClient.prompt().user(prompt).call().content(); return parseJsonArray(result); } /** * 將大模型返回的JSON數(shù)組字符串解析為ListString * 大模型返回格式示例[改寫版本1, 改寫版本2, 改寫版本3] */ private ListString parseJsonArray(String json) { if (json null || json.isBlank()) { return List.of(); } try { String cleaned json.trim(); if (cleaned.startsWith(json)) { cleaned cleaned.substring(7); } else if (cleaned.startsWith()) { cleaned cleaned.substring(3); } if (cleaned.endsWith()) { cleaned cleaned.substring(0, cleaned.length() - 3); } cleaned cleaned.trim(); return objectMapper.readValue(cleaned, new TypeReferenceListString() {}); } catch (Exception e) { return Arrays.stream(json.split(\n)) .map(String::trim) .filter(s - !s.isEmpty()) .map(s - s.replaceAll(^\\d[\\.\\)\u3001]\u0020*, )) .filter(s - !s.startsWith([)) .toList(); } } }4. 多查詢檢索流程GetMapping(/ask-multi-query) public FluxString askMultiQuery( RequestParam String message, RequestParam(defaultValue default) String conversationId) { // Step 1: 改寫問題生成3個版本 ListString queries queryRewriteService.rewrite(message, ); // Step 2: 分別檢索并去重 SetString seen new HashSet(); ListDocument allDocs new ArrayList(); for (String q : queries) { ListDocument docs knowledgeService.search(q, 5, 0.6); for (Document doc : docs) { if (seen.add(doc.getId())) { allDocs.add(doc); } } } // Step 3: 生成答案 String context buildContext(allDocs); return chatClient.prompt() .user(參考資料\n context \n\n問題 message) .stream().content(); }四、閑聊感知查詢改寫問題背景用戶可能輸入閑聊內容如 你好此時不應觸發(fā)RAG檢索流程避免浪費資源和產(chǎn)生無關檢索結果。public class ChitchatAwareQueryTransformer implements QueryTransformer { private final RewriteQueryTransformer delegate; public ChitchatAwareQueryTransformer(ChatClient.Builder chatClientBuilder) { this.delegate RewriteQueryTransformer.builder() .chatClientBuilder(chatClientBuilder) .targetSearchSystem(vector store) .build(); } Override public Query transform(Query query) { String text query.text(); if (isChitchat(text)) { return query; // 閑聊直接透傳 } return delegate.transform(query); // 知識性問題交給改寫器 } private boolean isChitchat(String text) { if (text null || text.isBlank()) return true; String trimmed text.trim(); if (trimmed.length() 6) return true; String lower trimmed.toLowerCase(); return lower.matches(.*(你好|您好|hi|hello|謝謝|感謝|再見|你是誰).*) !lower.matches(.*(什么是|怎么|如何|為什么|區(qū)別|原理|實現(xiàn)).*); } }集成到RAG AdvisorRetrievalAugmentationAdvisor ragAdvisor RetrievalAugmentationAdvisor.builder() .queryTransformers(new ChitchatAwareQueryTransformer(ChatClient.builder(chatModel))) .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .similarityThreshold(0.7) .topK(6) .build()) .build();五、查詢增強與空上下文處理ContextualQueryAugmenter負責將檢索到的文檔與用戶查詢拼裝為完整的 Prompt。allowEmptyContext 配置默認情況下無檢索結果時會拋出異常。設置 allowEmptyContext(true) 后無檢索結果時仍能正常生成回答。Bean public RetrievalAugmentationAdvisor ragAdvisor( VectorStore vectorStore) { return RetrievalAugmentationAdvisor.builder() .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .topK(5) .build()) .queryAugmenter(ContextualQueryAugmenter.builder() .allowEmptyContext(true) // 允許空上下文 .build()) .build(); }六、幻覺檢測與答案質量保障1. 基于引用驗證的幻覺檢測幻覺檢測的核心思路檢查模型生成的答案是否完全基于參考資料識別不受支持的說法。Service public class HallucinationDetector { private final ChatClient chatClient; private final ObjectMapper objectMapper; public HallucinationDetector(ChatClient.Builder builder) { this.chatClient builder .defaultSystem(你是一個嚴格的事實核查專家。) .build(); this.objectMapper new ObjectMapper(); } public MapString, Object detect(String answer, String context) { String prompt String.format( 請判斷以下答案是否完全基于參考資料得出。 參考資料%s 模型答案%s 請用JSON格式返回 {reliable: true/false, reason: 判斷理由, unsupportedClaims: [...]} , context, answer); String result chatClient.prompt().user(prompt).call().content(); return parseResult(result); } /** * 將大模型返回的幻覺檢測結果JSON字符串解析為Map */ private MapString, Object parseResult(String json) { MapString, Object defaultResult Map.of( reliable, true, reason, 解析失敗默認標記為可信 ); if (json null || json.isBlank()) { return defaultResult; } try { String cleaned json.trim(); if (cleaned.startsWith(json)) { cleaned cleaned.substring(7); } else if (cleaned.startsWith()) { cleaned cleaned.substring(3); } if (cleaned.endsWith()) { cleaned cleaned.substring(0, cleaned.length() - 3); } cleaned cleaned.trim(); JsonNode node objectMapper.readTree(cleaned); boolean reliable node.has(reliable) node.get(reliable).asBoolean(); String reason node.has(reason) ? node.get(reason).asText() : ; ListString unsupportedClaims new ArrayList(); if (node.has(unsupportedClaims) node.get(unsupportedClaims).isArray()) { for (JsonNode claim : node.get(unsupportedClaims)) { unsupportedClaims.add(claim.asText()); } } return Map.of(reliable, reliable, reason, reason, unsupportedClaims, unsupportedClaims); } catch (Exception e) { return defaultResult; } } }2. RAG流程集成GetMapping(/ask-safe) public MapString, Object askSafe( RequestParam String message, RequestParam(defaultValue default) String conversationId) { // Step 1: 檢索文檔 ListDocument docs vectorStore.similaritySearch( SearchRequest.builder().query(message).topK(5).build() ); // Step 2: 生成答案 String context buildContext(docs); String answer chatClient.prompt() .user(參考資料\n context \n\n問題 message) .call().content(); // Step 3: 幻覺檢測 MapString, Object result hallucinationDetector.detect(answer, context); // Step 4: 組裝返回 MapString, Object response new HashMap(); response.put(answer, answer); response.put(reliable, result.get(reliable)); response.put(reason, result.get(reason)); response.put(unsupportedClaims, result.get(unsupportedClaims)); return response; }3. System Prompt約束建議只基于提供的參考資料回答問題不要編造信息。如果參考資料中沒有相關內容請明確告知用戶。引用參考資料時標注來源。七、RAG Advisor完整優(yōu)化流程核心流程查詢改寫 → 檢索 → 重排 → 增強 → 生成完整配置示例Bean public ChatClient ragOptimizedChatClient( ChatClient.Builder builder, VectorStore vectorStore, ChatModel chatModel) { String promptTemplate 你是一個知識庫助手。請嚴格遵循以下規(guī)則 1. 只基于提供的參考資料回答問題不要編造信息 2. 如果參考資料中沒有相關內容請明確告知用戶 3. 引用參考資料時標注來源 參考資料 {context} 問題{question} ; RetrievalAugmentationAdvisor advisor RetrievalAugmentationAdvisor.builder() .queryTransformers(new ChitchatAwareQueryTransformer(builder)) .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .similarityThreshold(0.7) .topK(6) .build()) .documentPostProcessors(new DashScopeDocumentReranker(apiKey, qwen3-rerank, 3)) .queryAugmenter(ContextualQueryAugmenter.builder() .allowEmptyContext(true) .build()) .build(); return builder.defaultAdvisors(advisor).build(); }八、常用工具方法buildContext文檔列表拼接為上下文字符串將檢索到的 Document 列表拼接為一段完整的上下文文本。private String buildContext(ListDocument documents) { if (documents null || documents.isEmpty()) { return ; } StringBuilder sb new StringBuilder(); for (int i 0; i documents.size(); i) { Document doc documents.get(i); sb.append(--- 文檔).append(i 1).append( ---\n); Object fileName doc.getMetadata().get(fileName); if (fileName ! null) { sb.append(來源).append(fileName).append(\n); } sb.append(doc.getText()).append(\n\n); } return sb.toString(); }extractCitations從響應中提取引用來源解析 Spring AI ChatResponse 中的 citations 信息。/** * 從ChatResponse中提取引用的文檔信息 */ private ListMapString, Object extractCitations(ChatResponse response) { ListMapString, Object citations new ArrayList(); if (response null || response.getMetadata() null) { return citations; } MapString, Object metadata response.getMetadata(); Object docsObj metadata.get(documents); if (docsObj instanceof List?) { for (Object obj : (List?) docsObj) { if (obj instanceof Document doc) { MapString, Object citation new HashMap(); citation.put(content, doc.getText()); citation.put(metadata, doc.getMetadata()); citations.add(citation); } } } return citations; }總結結合混合檢索、Rerank重排序、Query改寫、閑聊感知、查詢增強和幻覺檢測可顯著提升知識庫問答系統(tǒng)的檢索效果與答案質量。建議在實際應用中分階段引入上述機制持續(xù)優(yōu)化用戶體驗和結果可靠性。 第二章：5、RAG進階優(yōu)化
一、混合檢索向量 關鍵詞1. 純向量檢索的局限語義相近但答案靠后向量相似度高不代表文檔真正包含答案。精確關鍵詞匹配弱用戶搜索特定術語時向量檢索可能漏掉精確匹配的文檔。多路召回更可靠結合向量檢索和關鍵詞檢索取并集后融合排序。2. 混合檢索的實現(xiàn)以PGVector為例-- PGVector 全文檢索 向量檢索 SQL SELECT * FROM knowledge_base WHERE tsv to_tsquery(chinese, :query) -- 關鍵詞匹配 ORDER BY ts_rank(tsv, to_tsquery(chinese, :query)) DESC, -- 關鍵詞相關性 embedding :query_vector -- 向量相似度 LIMIT :topK;3. Java端混合檢索邏輯核心思路向量檢索與關鍵詞檢索并行結果融合去重RRF排序。Service public class HybridSearchService { Resource private VectorStore vectorStore; Resource private JdbcTemplate jdbcTemplate; public ListDocument hybridSearch(String query, int topK) { // 1. 向量檢索 ListDocument vectorResults vectorStore.similaritySearch( SearchRequest.builder().query(query).topK(topK).build() ); // 2. 關鍵詞檢索PGVector全文搜索 String sql SELECT id, content, ts_rank(tsv, to_tsquery(chinese, ?)) AS rank FROM knowledge_base WHERE tsv to_tsquery(chinese, ?) ORDER BY rank DESC LIMIT ? ; ListDocument keywordResults jdbcTemplate.query(sql, ...); // 3. RRF融合排序Reciprocal Rank fusion return rrfMerge(vectorResults, keywordResults, topK); } /** * RRFReciprocal Rank Fusion融合排序 * 對兩個檢索結果按排名的倒數(shù)加權融合得到最終排序 */ private ListDocument rrfMerge( ListDocument vectorResults, ListDocument keywordResults, int topK) { double k 60.0; // RRF常量 MapString, Double scores new HashMap(); MapString, Document docMap new HashMap(); // 向量檢索結果按排名加權 for (int i 0; i vectorResults.size(); i) { String id vectorResults.get(i).getId(); double score 1.0 / (k i 1); scores.merge(id, score, Double::sum); docMap.put(id, vectorResults.get(i)); } // 關鍵詞檢索結果按排名加權 for (int i 0; i keywordResults.size(); i) { String id keywordResults.get(i).getId(); double score 1.0 / (k i 1); scores.merge(id, score, Double::sum); docMap.put(id, keywordResults.get(i)); } // 按融合分數(shù)降序取topK return scores.entrySet().stream() .sorted(Map.Entry.String, DoublecomparingByValue().reversed()) .limit(topK) .map(e - docMap.get(e.getKey())) .toList(); } }4. 集成到RAG流程GetMapping(/ask-hybrid) public String askHybrid(RequestParam String question) { // Step 1: 混合檢索 ListDocument docs hybridSearchService.hybridSearch(question, 10); // Step 2: 生成答案 String context buildContext(docs); return chatClient.prompt() .user(參考資料\n context \n\n問題 question) .call().content(); }二、Rerank重排序Rerank 是對檢索結果按查詢相關性進行二次精排的過程。粗檢索階段多召回topK20Rerank階段精排取Top5。1. DashScope Rerankqwen3-rerank通過阿里云 DashScope API 調用 qwen3-rerank 模型實現(xiàn)基于API的文檔重排序。Service public class DashScopeDocumentReranker implements DocumentPostProcessor { private final RestTemplate restTemplate; private final String apiKey; private final String modelName; private final int topN; public DashScopeDocumentReranker(String apiKey, String modelName, int topN) { this.restTemplate new RestTemplate(); this.apiKey apiKey; this.modelName modelName; this.topN topN; } Override public ListDocument process(Query query, ListDocument documents) { if (documents null || documents.size() 1) { return documents; } ListString texts documents.stream().map(Document::getText).toList(); MapString, Object body Map.of( model, modelName, query, query.text(), documents, texts, top_n, Math.min(topN, documents.size()) ); HttpHeaders headers new HttpHeaders(); headers.setContentType(MediaType.APPLICATION_JSON); headers.set(Authorization, Bearer apiKey); try { ResponseEntityRerankResponse response restTemplate.postForEntity( https://dashscope.aliyuncs.com/api/v1/services/rerank, new HttpEntity(body, headers), RerankResponse.class ); return reorder(documents, response.getBody()); } catch (Exception e) { log.warn(DashScope rerank調用失敗: {}, e.getMessage()); return documents; } } /** * 按rerank返回的相關性順序重排文檔結果已按relevance_score降序排列 */ private ListDocument reorder(ListDocument documents, RerankResponse response) { if (response null || response.results null || response.results.isEmpty()) { return documents; } ListDocument reranked new ArrayList(); for (RerankResult result : response.results) { if (result.index ! null result.index 0 result.index documents.size()) { reranked.add(documents.get(result.index)); } } return reranked.isEmpty() ? documents : reranked; } // qwen3-rerank API 響應結構 static class RerankResponse { public ListRerankResult results; } static class RerankResult { JsonProperty(relevance_score) public Double relevanceScore; public Integer index; } }2. 硅基流動 RerankBAAI/bge-reranker-v2-m3調用硅基流動平臺的 bge-reranker-v2-m3 模型進行文檔重排序。Service public class RerankService { private static final String API_URL https://api.siliconflow.cn/v1/rerank; private final RestTemplate restTemplate; private final ObjectMapper objectMapper; Value(${spring.ai.siliconflow-api-key}) private String apiKey; public RerankService(RestTemplateBuilder restTemplateBuilder) { this.restTemplate restTemplateBuilder .connectTimeout(Duration.ofSeconds(10)) .readTimeout(Duration.ofSeconds(30)) .build(); this.objectMapper new ObjectMapper(); } public ListDocument rerank(String query, ListDocument documents, int topN) { if (documents null || documents.isEmpty()) { return List.of(); } if (documents.size() 1) { return documents; } ListString texts documents.stream().map(Document::getText).toList(); MapString, Object body Map.of( model, BAAI/bge-reranker-v2-m3, query, query, documents, texts, top_n, Math.min(topN, documents.size()) ); try { HttpHeaders headers new HttpHeaders(); headers.setContentType(MediaType.APPLICATION_JSON); headers.set(Authorization, Bearer apiKey); HttpEntityMapString, Object request new HttpEntity(body, headers); ResponseEntityMap response restTemplate.postForEntity( API_URL, request, Map.class ); SuppressWarnings(unchecked) ListMapString, Object results (ListMapString, Object) response.getBody().get(results); if (results null || results.isEmpty()) { return documents; } return results.stream().map(r - { int index (int) r.get(index); Document doc documents.get(index); doc.getMetadata().put(rerankScore, r.get(relevance_score)); return doc; }).toList(); } catch (Exception e) { log.warn(硅基流動Rerank調用失敗: {}, e.getMessage()); return documents.stream().limit(topN).toList(); } } }3. LLM文檔重排器自定義實現(xiàn)Spring AI 2.0.1 沒有內置 reranker可以用 LLM 對檢索結果按相關性重新排序。/** * 基于LLM的文檔重排器實現(xiàn)DocumentPostProcessor接口 */ public class LlmDocumentReranker implements DocumentPostProcessor { private static final Pattern INDEX_PATTERN Pattern.compile(\\d); private final ChatClient chatClient; private final int maxDocPreviewLength; public LlmDocumentReranker(ChatClient.Builder chatClientBuilder, int maxDocPreviewLength) { this.chatClient chatClientBuilder.build(); this.maxDocPreviewLength maxDocPreviewLength; } Override public ListDocument process(Query query, ListDocument documents) { if (documents null || documents.size() 1) { return documents; } StringBuilder docList new StringBuilder(); for (int i 0; i documents.size(); i) { String text documents.get(i).getText(); if (text null) { text ; } if (text.length() maxDocPreviewLength) { text text.substring(0, maxDocPreviewLength) ...; } docList.append([).append(i).append(] ).append(text).append(\n\n); } String prompt 你是一個文檔相關性排序助手。請根據(jù)文檔與查詢的相關程度重新排序。 查詢%s 文檔列表 %s 請只返回文檔編號的排序結果用英文逗號分隔。 例如2,0,3,1 .formatted(query.text(), docList); String result chatClient.prompt().user(prompt).call().content(); return reorder(documents, result); } /** * 解析LLM返回的編號序列按順序重排文檔。解析失敗時返回原順序。 */ private ListDocument reorder(ListDocument documents, String llmOutput) { if (llmOutput null || llmOutput.isBlank()) { return documents; } Matcher matcher INDEX_PATTERN.matcher(llmOutput); ListInteger indices new ArrayList(); while (matcher.find()) { int idx Integer.parseInt(matcher.group()); if (idx 0 idx documents.size() !indices.contains(idx)) { indices.add(idx); } } if (indices.size() ! documents.size()) { return documents; } ListDocument reranked new ArrayList(documents.size()); for (int idx : indices) { reranked.add(documents.get(idx)); } return reranked; } }4. Rerank集成方式方式一手動編排GetMapping(/ask-rerank) public String askWithRerank(RequestParam String question) { // Step 1: 粗檢索多召回 ListDocument candidates vectorStore.similaritySearch( SearchRequest.builder().query(question).topK(20).build() ); // Step 2: Rerank精排取Top5 ListDocument reranked rerankService.rerank(question, candidates, 5); // Step 3: 生成答案 String context buildContext(reranked); return chatClient.prompt() .user(參考資料\n context \n\n問題 question) .call().content(); }方式二Advisor集成RetrievalAugmentationAdvisor ragAdvisor RetrievalAugmentationAdvisor.builder() .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .similarityThreshold(0.7) .topK(6) .build()) .documentPostProcessors(new DashScopeDocumentReranker(apiKey, qwen3-rerank, 3)) .build();三、Query改寫與多查詢擴展1. Query改寫的必要性用戶提問可能口語化、模糊或含指代詞直接用于檢索效果差。通過LLM改寫為更適合知識庫檢索的形式提升召回率。2. RewriteQueryTransformerSpring AI內置Spring AI 2.0 提供了內置的查詢改寫器可直接集成到 Advisor 中。Bean public RetrievalAugmentationAdvisor ragAdvisorWithRewrite( ChatClient.Builder chatClientBuilder, VectorStore vectorStore) { RewriteQueryTransformer rewriteTransformer RewriteQueryTransformer.builder() .chatClientBuilder(chatClientBuilder) .targetSearchSystem(vector store) .build(); return RetrievalAugmentationAdvisor.builder() .queryTransformers(rewriteTransformer) .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .topK(5) .build()) .build(); }3. 自定義多查詢改寫服務生成多個改寫版本分別檢索后取并集進一步提升召回率。Service public class QueryRewriteService { Resource private ChatClient ragManualChatClient; private final ObjectMapper objectMapper; public QueryRewriteService() { this.objectMapper new ObjectMapper(); } public ListString rewrite(String originalQuery, String conversationHistory) { String prompt String.format( 你是一個搜索查詢優(yōu)化專家。請將用戶的問題改寫為更適合在知識庫中檢索的形式。 要求 1. 生成3個不同角度的改寫版本 2. 用JSON數(shù)組格式返回 3. 保留原始問題的核心意圖 4. 如果有對話歷史結合上下文補全指代 對話歷史%s 用戶問題%s 請直接返回JSON數(shù)組不要包含其他文字。 , conversationHistory, originalQuery); String result ragManualChatClient.prompt().user(prompt).call().content(); return parseJsonArray(result); } /** * 將大模型返回的JSON數(shù)組字符串解析為ListString * 大模型返回格式示例[改寫版本1, 改寫版本2, 改寫版本3] */ private ListString parseJsonArray(String json) { if (json null || json.isBlank()) { return List.of(); } try { String cleaned json.trim(); if (cleaned.startsWith(json)) { cleaned cleaned.substring(7); } else if (cleaned.startsWith()) { cleaned cleaned.substring(3); } if (cleaned.endsWith()) { cleaned cleaned.substring(0, cleaned.length() - 3); } cleaned cleaned.trim(); return objectMapper.readValue(cleaned, new TypeReferenceListString() {}); } catch (Exception e) { return Arrays.stream(json.split(\n)) .map(String::trim) .filter(s - !s.isEmpty()) .map(s - s.replaceAll(^\\d[\\.\\)\u3001]\u0020*, )) .filter(s - !s.startsWith([)) .toList(); } } }4. 多查詢檢索流程GetMapping(/ask-multi-query) public FluxString askMultiQuery( RequestParam String message, RequestParam(defaultValue default) String conversationId) { // Step 1: 改寫問題生成3個版本 ListString queries queryRewriteService.rewrite(message, ); // Step 2: 分別檢索并去重 SetString seen new HashSet(); ListDocument allDocs new ArrayList(); for (String q : queries) { ListDocument docs knowledgeService.search(q, 5, 0.6); for (Document doc : docs) { if (seen.add(doc.getId())) { allDocs.add(doc); } } } // Step 3: 生成答案 String context buildContext(allDocs); return chatClient.prompt() .user(參考資料\n context \n\n問題 message) .stream().content(); }四、閑聊感知查詢改寫問題背景用戶可能輸入閑聊內容如 你好此時不應觸發(fā)RAG檢索流程避免浪費資源和產(chǎn)生無關檢索結果。public class ChitchatAwareQueryTransformer implements QueryTransformer { private final RewriteQueryTransformer delegate; public ChitchatAwareQueryTransformer(ChatClient.Builder chatClientBuilder) { this.delegate RewriteQueryTransformer.builder() .chatClientBuilder(chatClientBuilder) .targetSearchSystem(vector store) .build(); } Override public Query transform(Query query) { String text query.text(); if (isChitchat(text)) { return query; // 閑聊直接透傳 } return delegate.transform(query); // 知識性問題交給改寫器 } private boolean isChitchat(String text) { if (text null || text.isBlank()) return true; String trimmed text.trim(); if (trimmed.length() 6) return true; String lower trimmed.toLowerCase(); return lower.matches(.*(你好|您好|hi|hello|謝謝|感謝|再見|你是誰).*) !lower.matches(.*(什么是|怎么|如何|為什么|區(qū)別|原理|實現(xiàn)).*); } }集成到RAG AdvisorRetrievalAugmentationAdvisor ragAdvisor RetrievalAugmentationAdvisor.builder() .queryTransformers(new ChitchatAwareQueryTransformer(ChatClient.builder(chatModel))) .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .similarityThreshold(0.7) .topK(6) .build()) .build();五、查詢增強與空上下文處理ContextualQueryAugmenter負責將檢索到的文檔與用戶查詢拼裝為完整的 Prompt。allowEmptyContext 配置默認情況下無檢索結果時會拋出異常。設置 allowEmptyContext(true) 后無檢索結果時仍能正常生成回答。Bean public RetrievalAugmentationAdvisor ragAdvisor( VectorStore vectorStore) { return RetrievalAugmentationAdvisor.builder() .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .topK(5) .build()) .queryAugmenter(ContextualQueryAugmenter.builder() .allowEmptyContext(true) // 允許空上下文 .build()) .build(); }六、幻覺檢測與答案質量保障1. 基于引用驗證的幻覺檢測幻覺檢測的核心思路檢查模型生成的答案是否完全基于參考資料識別不受支持的說法。Service public class HallucinationDetector { private final ChatClient chatClient; private final ObjectMapper objectMapper; public HallucinationDetector(ChatClient.Builder builder) { this.chatClient builder .defaultSystem(你是一個嚴格的事實核查專家。) .build(); this.objectMapper new ObjectMapper(); } public MapString, Object detect(String answer, String context) { String prompt String.format( 請判斷以下答案是否完全基于參考資料得出。 參考資料%s 模型答案%s 請用JSON格式返回 {reliable: true/false, reason: 判斷理由, unsupportedClaims: [...]} , context, answer); String result chatClient.prompt().user(prompt).call().content(); return parseResult(result); } /** * 將大模型返回的幻覺檢測結果JSON字符串解析為Map */ private MapString, Object parseResult(String json) { MapString, Object defaultResult Map.of( reliable, true, reason, 解析失敗默認標記為可信 ); if (json null || json.isBlank()) { return defaultResult; } try { String cleaned json.trim(); if (cleaned.startsWith(json)) { cleaned cleaned.substring(7); } else if (cleaned.startsWith()) { cleaned cleaned.substring(3); } if (cleaned.endsWith()) { cleaned cleaned.substring(0, cleaned.length() - 3); } cleaned cleaned.trim(); JsonNode node objectMapper.readTree(cleaned); boolean reliable node.has(reliable) node.get(reliable).asBoolean(); String reason node.has(reason) ? node.get(reason).asText() : ; ListString unsupportedClaims new ArrayList(); if (node.has(unsupportedClaims) node.get(unsupportedClaims).isArray()) { for (JsonNode claim : node.get(unsupportedClaims)) { unsupportedClaims.add(claim.asText()); } } return Map.of(reliable, reliable, reason, reason, unsupportedClaims, unsupportedClaims); } catch (Exception e) { return defaultResult; } } }2. RAG流程集成GetMapping(/ask-safe) public MapString, Object askSafe( RequestParam String message, RequestParam(defaultValue default) String conversationId) { // Step 1: 檢索文檔 ListDocument docs vectorStore.similaritySearch( SearchRequest.builder().query(message).topK(5).build() ); // Step 2: 生成答案 String context buildContext(docs); String answer chatClient.prompt() .user(參考資料\n context \n\n問題 message) .call().content(); // Step 3: 幻覺檢測 MapString, Object result hallucinationDetector.detect(answer, context); // Step 4: 組裝返回 MapString, Object response new HashMap(); response.put(answer, answer); response.put(reliable, result.get(reliable)); response.put(reason, result.get(reason)); response.put(unsupportedClaims, result.get(unsupportedClaims)); return response; }3. System Prompt約束建議只基于提供的參考資料回答問題不要編造信息。如果參考資料中沒有相關內容請明確告知用戶。引用參考資料時標注來源。七、RAG Advisor完整優(yōu)化流程核心流程查詢改寫 → 檢索 → 重排 → 增強 → 生成完整配置示例Bean public ChatClient ragOptimizedChatClient( ChatClient.Builder builder, VectorStore vectorStore, ChatModel chatModel) { String promptTemplate 你是一個知識庫助手。請嚴格遵循以下規(guī)則 1. 只基于提供的參考資料回答問題不要編造信息 2. 如果參考資料中沒有相關內容請明確告知用戶 3. 引用參考資料時標注來源 參考資料 {context} 問題{question} ; RetrievalAugmentationAdvisor advisor RetrievalAugmentationAdvisor.builder() .queryTransformers(new ChitchatAwareQueryTransformer(builder)) .documentRetriever(VectorStoreDocumentRetriever.builder() .vectorStore(vectorStore) .similarityThreshold(0.7) .topK(6) .build()) .documentPostProcessors(new DashScopeDocumentReranker(apiKey, qwen3-rerank, 3)) .queryAugmenter(ContextualQueryAugmenter.builder() .allowEmptyContext(true) .build()) .build(); return builder.defaultAdvisors(advisor).build(); }八、常用工具方法buildContext文檔列表拼接為上下文字符串將檢索到的 Document 列表拼接為一段完整的上下文文本。private String buildContext(ListDocument documents) { if (documents null || documents.isEmpty()) { return ; } StringBuilder sb new StringBuilder(); for (int i 0; i documents.size(); i) { Document doc documents.get(i); sb.append(--- 文檔).append(i 1).append( ---\n); Object fileName doc.getMetadata().get(fileName); if (fileName ! null) { sb.append(來源).append(fileName).append(\n); } sb.append(doc.getText()).append(\n\n); } return sb.toString(); }extractCitations從響應中提取引用來源解析 Spring AI ChatResponse 中的 citations 信息。/** * 從ChatResponse中提取引用的文檔信息 */ private ListMapString, Object extractCitations(ChatResponse response) { ListMapString, Object citations new ArrayList(); if (response null || response.getMetadata() null) { return citations; } MapString, Object metadata response.getMetadata(); Object docsObj metadata.get(documents); if (docsObj instanceof List?) { for (Object obj : (List?) docsObj) { if (obj instanceof Document doc) { MapString, Object citation new HashMap(); citation.put(content, doc.getText()); citation.put(metadata, doc.getMetadata()); citations.add(citation); } } } return citations; }總結結合混合檢索、Rerank重排序、Query改寫、閑聊感知、查詢增強和幻覺檢測可顯著提升知識庫問答系統(tǒng)的檢索效果與答案質量。建議在實際應用中分階段引入上述機制持續(xù)優(yōu)化用戶體驗和結果可靠性。