Commit ed136ff02224f9763c373d7c62b12e7a9cdeef98
1 parent
4af529cb
feat:【ai】增加智能文档切片策略,支持自动识别 Markdown QA 和语义化切分「代码优化」
Showing
5 changed files
with
88 additions
and
107 deletions
pom.xml
| @@ -23,7 +23,7 @@ | @@ -23,7 +23,7 @@ | ||
| 23 | <!-- <module>yudao-module-mall</module>--> | 23 | <!-- <module>yudao-module-mall</module>--> |
| 24 | <!-- <module>yudao-module-crm</module>--> | 24 | <!-- <module>yudao-module-crm</module>--> |
| 25 | <!-- <module>yudao-module-erp</module>--> | 25 | <!-- <module>yudao-module-erp</module>--> |
| 26 | - <module>yudao-module-ai</module> | 26 | +<!-- <module>yudao-module-ai</module>--> |
| 27 | <!-- <module>yudao-module-iot</module>--> | 27 | <!-- <module>yudao-module-iot</module>--> |
| 28 | </modules> | 28 | </modules> |
| 29 | 29 |
yudao-module-ai/src/main/java/cn/iocoder/yudao/module/ai/enums/AiDocumentSplitStrategyEnum.java
| @@ -50,16 +50,4 @@ public enum AiDocumentSplitStrategyEnum { | @@ -50,16 +50,4 @@ public enum AiDocumentSplitStrategyEnum { | ||
| 50 | */ | 50 | */ |
| 51 | private final String name; | 51 | private final String name; |
| 52 | 52 | ||
| 53 | - /** | ||
| 54 | - * 根据代码获取枚举 | ||
| 55 | - */ | ||
| 56 | - public static AiDocumentSplitStrategyEnum fromCode(String code) { | ||
| 57 | - for (AiDocumentSplitStrategyEnum strategy : values()) { | ||
| 58 | - if (strategy.getCode().equals(code)) { | ||
| 59 | - return strategy; | ||
| 60 | - } | ||
| 61 | - } | ||
| 62 | - return AUTO; // 默认返回自动识别 | ||
| 63 | - } | ||
| 64 | - | ||
| 65 | } | 53 | } |
yudao-module-ai/src/main/java/cn/iocoder/yudao/module/ai/service/knowledge/AiKnowledgeSegmentServiceImpl.java
| @@ -107,11 +107,8 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | @@ -107,11 +107,8 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | ||
| 107 | if (StrUtil.isEmpty(segment.getText())) { | 107 | if (StrUtil.isEmpty(segment.getText())) { |
| 108 | return null; | 108 | return null; |
| 109 | } | 109 | } |
| 110 | - return new AiKnowledgeSegmentDO() | ||
| 111 | - .setKnowledgeId(documentDO.getKnowledgeId()) | ||
| 112 | - .setDocumentId(documentId) | ||
| 113 | - .setContent(segment.getText()) | ||
| 114 | - .setContentLength(segment.getText().length()) | 110 | + return new AiKnowledgeSegmentDO().setKnowledgeId(documentDO.getKnowledgeId()).setDocumentId(documentId) |
| 111 | + .setContent(segment.getText()).setContentLength(segment.getText().length()) | ||
| 115 | .setVectorId(AiKnowledgeSegmentDO.VECTOR_ID_EMPTY) | 112 | .setVectorId(AiKnowledgeSegmentDO.VECTOR_ID_EMPTY) |
| 116 | .setTokens(tokenCountEstimator.estimate(segment.getText())) | 113 | .setTokens(tokenCountEstimator.estimate(segment.getText())) |
| 117 | .setStatus(CommonStatusEnum.ENABLE.getStatus()); | 114 | .setStatus(CommonStatusEnum.ENABLE.getStatus()); |
| @@ -302,13 +299,12 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | @@ -302,13 +299,12 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | ||
| 302 | // 1. 读取 URL 内容 | 299 | // 1. 读取 URL 内容 |
| 303 | String content = knowledgeDocumentService.readUrl(url); | 300 | String content = knowledgeDocumentService.readUrl(url); |
| 304 | 301 | ||
| 305 | - // 2. 自动检测文档类型并选择策略 | 302 | + // 2.1 自动检测文档类型并选择策略 |
| 306 | AiDocumentSplitStrategyEnum strategy = detectDocumentStrategy(content, url); | 303 | AiDocumentSplitStrategyEnum strategy = detectDocumentStrategy(content, url); |
| 307 | - | ||
| 308 | - // 3. 文档切片 | 304 | + // 2.2 文档切片 |
| 309 | List<Document> documentSegments = splitContentByStrategy(content, segmentMaxTokens, strategy, url); | 305 | List<Document> documentSegments = splitContentByStrategy(content, segmentMaxTokens, strategy, url); |
| 310 | 306 | ||
| 311 | - // 4. 转换为段落对象 | 307 | + // 3. 转换为段落对象 |
| 312 | return convertList(documentSegments, segment -> { | 308 | return convertList(documentSegments, segment -> { |
| 313 | if (StrUtil.isEmpty(segment.getText())) { | 309 | if (StrUtil.isEmpty(segment.getText())) { |
| 314 | return null; | 310 | return null; |
| @@ -352,6 +348,7 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | @@ -352,6 +348,7 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | ||
| 352 | * @param url 文档 URL(用于自动检测文件类型) | 348 | * @param url 文档 URL(用于自动检测文件类型) |
| 353 | * @return 切片后的文档列表 | 349 | * @return 切片后的文档列表 |
| 354 | */ | 350 | */ |
| 351 | + @SuppressWarnings("EnhancedSwitchMigration") | ||
| 355 | private List<Document> splitContentByStrategy(String content, Integer segmentMaxTokens, | 352 | private List<Document> splitContentByStrategy(String content, Integer segmentMaxTokens, |
| 356 | AiDocumentSplitStrategyEnum strategy, String url) { | 353 | AiDocumentSplitStrategyEnum strategy, String url) { |
| 357 | // 自动检测策略 | 354 | // 自动检测策略 |
| @@ -359,7 +356,7 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | @@ -359,7 +356,7 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | ||
| 359 | strategy = detectDocumentStrategy(content, url); | 356 | strategy = detectDocumentStrategy(content, url); |
| 360 | log.info("[splitContentByStrategy][自动检测到文档策略: {}]", strategy.getName()); | 357 | log.info("[splitContentByStrategy][自动检测到文档策略: {}]", strategy.getName()); |
| 361 | } | 358 | } |
| 362 | - | 359 | + // 根据策略切分 |
| 363 | TextSplitter textSplitter; | 360 | TextSplitter textSplitter; |
| 364 | switch (strategy) { | 361 | switch (strategy) { |
| 365 | case MARKDOWN_QA: | 362 | case MARKDOWN_QA: |
| @@ -376,7 +373,7 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | @@ -376,7 +373,7 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | ||
| 376 | textSplitter = buildTokenTextSplitter(segmentMaxTokens); | 373 | textSplitter = buildTokenTextSplitter(segmentMaxTokens); |
| 377 | break; | 374 | break; |
| 378 | } | 375 | } |
| 379 | - | 376 | + // 执行切分 |
| 380 | return textSplitter.apply(Collections.singletonList(new Document(content))); | 377 | return textSplitter.apply(Collections.singletonList(new Document(content))); |
| 381 | } | 378 | } |
| 382 | 379 | ||
| @@ -391,17 +388,14 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | @@ -391,17 +388,14 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | ||
| 391 | if (StrUtil.isEmpty(content)) { | 388 | if (StrUtil.isEmpty(content)) { |
| 392 | return AiDocumentSplitStrategyEnum.TOKEN; | 389 | return AiDocumentSplitStrategyEnum.TOKEN; |
| 393 | } | 390 | } |
| 394 | - | ||
| 395 | // 1. 检测 Markdown QA 格式 | 391 | // 1. 检测 Markdown QA 格式 |
| 396 | if (isMarkdownQaFormat(content, url)) { | 392 | if (isMarkdownQaFormat(content, url)) { |
| 397 | return AiDocumentSplitStrategyEnum.MARKDOWN_QA; | 393 | return AiDocumentSplitStrategyEnum.MARKDOWN_QA; |
| 398 | } | 394 | } |
| 399 | - | ||
| 400 | // 2. 检测普通 Markdown 文档 | 395 | // 2. 检测普通 Markdown 文档 |
| 401 | if (isMarkdownDocument(url)) { | 396 | if (isMarkdownDocument(url)) { |
| 402 | return AiDocumentSplitStrategyEnum.SEMANTIC; | 397 | return AiDocumentSplitStrategyEnum.SEMANTIC; |
| 403 | } | 398 | } |
| 404 | - | ||
| 405 | // 3. 默认使用语义切分(比 Token 切分更智能) | 399 | // 3. 默认使用语义切分(比 Token 切分更智能) |
| 406 | return AiDocumentSplitStrategyEnum.SEMANTIC; | 400 | return AiDocumentSplitStrategyEnum.SEMANTIC; |
| 407 | } | 401 | } |
| @@ -421,16 +415,14 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | @@ -421,16 +415,14 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | ||
| 421 | .filter(line -> line.trim().startsWith("## ")) | 415 | .filter(line -> line.trim().startsWith("## ")) |
| 422 | .count(); | 416 | .count(); |
| 423 | 417 | ||
| 424 | - // 至少包含 2 个二级标题才认为是 QA 格式 | 418 | + // 要求一:至少包含 2 个二级标题才认为是 QA 格式 |
| 425 | if (h2Count < 2) { | 419 | if (h2Count < 2) { |
| 426 | return false; | 420 | return false; |
| 427 | } | 421 | } |
| 428 | 422 | ||
| 429 | - // 检查标题占比(QA 文档标题行数相对较多) | 423 | + // 要求二:检查标题占比(QA 文档标题行数相对较多),如果二级标题占比超过 10%,认为是 QA 格式 |
| 430 | long totalLines = content.lines().count(); | 424 | long totalLines = content.lines().count(); |
| 431 | double h2Ratio = (double) h2Count / totalLines; | 425 | double h2Ratio = (double) h2Count / totalLines; |
| 432 | - | ||
| 433 | - // 如果二级标题占比超过 10%,认为是 QA 格式 | ||
| 434 | return h2Ratio > 0.1; | 426 | return h2Ratio > 0.1; |
| 435 | } | 427 | } |
| 436 | 428 | ||
| @@ -438,7 +430,7 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | @@ -438,7 +430,7 @@ public class AiKnowledgeSegmentServiceImpl implements AiKnowledgeSegmentService | ||
| 438 | * 检测是否为 Markdown 文档 | 430 | * 检测是否为 Markdown 文档 |
| 439 | */ | 431 | */ |
| 440 | private boolean isMarkdownDocument(String url) { | 432 | private boolean isMarkdownDocument(String url) { |
| 441 | - return StrUtil.isNotEmpty(url) && url.toLowerCase().endsWith(".md"); | 433 | + return StrUtil.endWithAnyIgnoreCase(url, ".md", ".markdown"); |
| 442 | } | 434 | } |
| 443 | 435 | ||
| 444 | /** | 436 | /** |
yudao-module-ai/src/main/java/cn/iocoder/yudao/module/ai/service/knowledge/splitter/MarkdownQaSplitter.java
| 1 | package cn.iocoder.yudao.module.ai.service.knowledge.splitter; | 1 | package cn.iocoder.yudao.module.ai.service.knowledge.splitter; |
| 2 | 2 | ||
| 3 | +import cn.hutool.core.collection.CollUtil; | ||
| 3 | import cn.hutool.core.util.StrUtil; | 4 | import cn.hutool.core.util.StrUtil; |
| 5 | +import lombok.AllArgsConstructor; | ||
| 4 | import lombok.extern.slf4j.Slf4j; | 6 | import lombok.extern.slf4j.Slf4j; |
| 5 | import org.springframework.ai.transformer.splitter.TextSplitter; | 7 | import org.springframework.ai.transformer.splitter.TextSplitter; |
| 6 | 8 | ||
| @@ -24,6 +26,7 @@ import java.util.regex.Pattern; | @@ -24,6 +26,7 @@ import java.util.regex.Pattern; | ||
| 24 | * @author runzhen | 26 | * @author runzhen |
| 25 | */ | 27 | */ |
| 26 | @Slf4j | 28 | @Slf4j |
| 29 | +@SuppressWarnings("SizeReplaceableByIsEmpty") | ||
| 27 | public class MarkdownQaSplitter extends TextSplitter { | 30 | public class MarkdownQaSplitter extends TextSplitter { |
| 28 | 31 | ||
| 29 | /** | 32 | /** |
| @@ -62,41 +65,38 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -62,41 +65,38 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 62 | return Collections.emptyList(); | 65 | return Collections.emptyList(); |
| 63 | } | 66 | } |
| 64 | 67 | ||
| 65 | - List<String> result = new ArrayList<>(); | ||
| 66 | - | ||
| 67 | // 解析 QA 对 | 68 | // 解析 QA 对 |
| 68 | List<QaPair> qaPairs = parseQaPairs(text); | 69 | List<QaPair> qaPairs = parseQaPairs(text); |
| 69 | - | ||
| 70 | - if (qaPairs.isEmpty()) { | 70 | + if (CollUtil.isEmpty(qaPairs)) { |
| 71 | // 如果没有识别到 QA 格式,按段落切分 | 71 | // 如果没有识别到 QA 格式,按段落切分 |
| 72 | return fallbackSplit(text); | 72 | return fallbackSplit(text); |
| 73 | } | 73 | } |
| 74 | 74 | ||
| 75 | // 处理每个 QA 对 | 75 | // 处理每个 QA 对 |
| 76 | + List<String> result = new ArrayList<>(); | ||
| 76 | for (QaPair qaPair : qaPairs) { | 77 | for (QaPair qaPair : qaPairs) { |
| 77 | result.addAll(splitQaPair(qaPair)); | 78 | result.addAll(splitQaPair(qaPair)); |
| 78 | } | 79 | } |
| 79 | - | ||
| 80 | return result; | 80 | return result; |
| 81 | } | 81 | } |
| 82 | 82 | ||
| 83 | /** | 83 | /** |
| 84 | * 解析 Markdown QA 对 | 84 | * 解析 Markdown QA 对 |
| 85 | + * | ||
| 86 | + * @param content 文本内容 | ||
| 87 | + * @return QA 对列表 | ||
| 85 | */ | 88 | */ |
| 86 | private List<QaPair> parseQaPairs(String content) { | 89 | private List<QaPair> parseQaPairs(String content) { |
| 90 | + // 找到所有二级标题位置 | ||
| 87 | List<QaPair> qaPairs = new ArrayList<>(); | 91 | List<QaPair> qaPairs = new ArrayList<>(); |
| 88 | - Matcher matcher = H2_PATTERN.matcher(content); | ||
| 89 | - | ||
| 90 | List<Integer> headingPositions = new ArrayList<>(); | 92 | List<Integer> headingPositions = new ArrayList<>(); |
| 91 | List<String> questions = new ArrayList<>(); | 93 | List<String> questions = new ArrayList<>(); |
| 92 | - | ||
| 93 | - // 找到所有二级标题位置 | 94 | + Matcher matcher = H2_PATTERN.matcher(content); |
| 94 | while (matcher.find()) { | 95 | while (matcher.find()) { |
| 95 | headingPositions.add(matcher.start()); | 96 | headingPositions.add(matcher.start()); |
| 96 | questions.add(matcher.group(1).trim()); | 97 | questions.add(matcher.group(1).trim()); |
| 97 | } | 98 | } |
| 98 | - | ||
| 99 | - if (headingPositions.isEmpty()) { | 99 | + if (CollUtil.isEmpty(headingPositions)) { |
| 100 | return qaPairs; | 100 | return qaPairs; |
| 101 | } | 101 | } |
| 102 | 102 | ||
| @@ -106,55 +106,51 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -106,55 +106,51 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 106 | int end = (i + 1 < headingPositions.size()) | 106 | int end = (i + 1 < headingPositions.size()) |
| 107 | ? headingPositions.get(i + 1) | 107 | ? headingPositions.get(i + 1) |
| 108 | : content.length(); | 108 | : content.length(); |
| 109 | - | ||
| 110 | String qaText = content.substring(start, end).trim(); | 109 | String qaText = content.substring(start, end).trim(); |
| 111 | String question = questions.get(i); | 110 | String question = questions.get(i); |
| 112 | - | ||
| 113 | // 提取答案部分(去掉问题标题) | 111 | // 提取答案部分(去掉问题标题) |
| 114 | String answer = qaText.substring(qaText.indexOf('\n') + 1).trim(); | 112 | String answer = qaText.substring(qaText.indexOf('\n') + 1).trim(); |
| 115 | - | ||
| 116 | qaPairs.add(new QaPair(question, answer, qaText)); | 113 | qaPairs.add(new QaPair(question, answer, qaText)); |
| 117 | } | 114 | } |
| 118 | - | ||
| 119 | return qaPairs; | 115 | return qaPairs; |
| 120 | } | 116 | } |
| 121 | 117 | ||
| 122 | /** | 118 | /** |
| 123 | * 切分单个 QA 对 | 119 | * 切分单个 QA 对 |
| 120 | + * | ||
| 121 | + * @param qaPair QA 对 | ||
| 122 | + * @return 切分后的文本片段列表 | ||
| 124 | */ | 123 | */ |
| 125 | private List<String> splitQaPair(QaPair qaPair) { | 124 | private List<String> splitQaPair(QaPair qaPair) { |
| 125 | + // 如果整个 QA 对不超过限制,保持完整 | ||
| 126 | List<String> chunks = new ArrayList<>(); | 126 | List<String> chunks = new ArrayList<>(); |
| 127 | - | ||
| 128 | String fullQa = qaPair.fullText; | 127 | String fullQa = qaPair.fullText; |
| 129 | int qaTokens = tokenEstimator.estimate(fullQa); | 128 | int qaTokens = tokenEstimator.estimate(fullQa); |
| 130 | - | ||
| 131 | - // 如果整个 QA 对不超过限制,保持完整 | ||
| 132 | if (qaTokens <= chunkSize) { | 129 | if (qaTokens <= chunkSize) { |
| 133 | chunks.add(fullQa); | 130 | chunks.add(fullQa); |
| 134 | return chunks; | 131 | return chunks; |
| 135 | } | 132 | } |
| 136 | 133 | ||
| 137 | // 长答案需要切分 | 134 | // 长答案需要切分 |
| 138 | - log.debug("QA 对超过 Token 限制 ({} > {}),开始智能切分: {}", | ||
| 139 | - qaTokens, chunkSize, qaPair.question); | ||
| 140 | - | 135 | + log.debug("QA 对超过 Token 限制 ({} > {}),开始智能切分: {}", qaTokens, chunkSize, qaPair.question); |
| 141 | List<String> answerChunks = splitLongAnswer(qaPair.answer, qaPair.question); | 136 | List<String> answerChunks = splitLongAnswer(qaPair.answer, qaPair.question); |
| 142 | - | ||
| 143 | for (String answerChunk : answerChunks) { | 137 | for (String answerChunk : answerChunks) { |
| 144 | // 每个片段都包含完整问题 | 138 | // 每个片段都包含完整问题 |
| 145 | String chunkText = "## " + qaPair.question + "\n" + answerChunk; | 139 | String chunkText = "## " + qaPair.question + "\n" + answerChunk; |
| 146 | chunks.add(chunkText); | 140 | chunks.add(chunkText); |
| 147 | } | 141 | } |
| 148 | - | ||
| 149 | return chunks; | 142 | return chunks; |
| 150 | } | 143 | } |
| 151 | 144 | ||
| 152 | /** | 145 | /** |
| 153 | * 切分长答案 | 146 | * 切分长答案 |
| 147 | + * | ||
| 148 | + * @param answer 答案文本 | ||
| 149 | + * @param question 问题文本 | ||
| 150 | + * @return 切分后的答案片段列表 | ||
| 154 | */ | 151 | */ |
| 155 | private List<String> splitLongAnswer(String answer, String question) { | 152 | private List<String> splitLongAnswer(String answer, String question) { |
| 156 | List<String> chunks = new ArrayList<>(); | 153 | List<String> chunks = new ArrayList<>(); |
| 157 | - | ||
| 158 | // 预留问题的 Token 空间 | 154 | // 预留问题的 Token 空间 |
| 159 | String questionHeader = "## " + question + "\n"; | 155 | String questionHeader = "## " + question + "\n"; |
| 160 | int questionTokens = tokenEstimator.estimate(questionHeader); | 156 | int questionTokens = tokenEstimator.estimate(questionHeader); |
| @@ -162,17 +158,13 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -162,17 +158,13 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 162 | 158 | ||
| 163 | // 先按段落切分 | 159 | // 先按段落切分 |
| 164 | String[] paragraphs = answer.split(PARAGRAPH_SEPARATOR); | 160 | String[] paragraphs = answer.split(PARAGRAPH_SEPARATOR); |
| 165 | - | ||
| 166 | StringBuilder currentChunk = new StringBuilder(); | 161 | StringBuilder currentChunk = new StringBuilder(); |
| 167 | int currentTokens = 0; | 162 | int currentTokens = 0; |
| 168 | - | ||
| 169 | for (String paragraph : paragraphs) { | 163 | for (String paragraph : paragraphs) { |
| 170 | if (StrUtil.isEmpty(paragraph)) { | 164 | if (StrUtil.isEmpty(paragraph)) { |
| 171 | continue; | 165 | continue; |
| 172 | } | 166 | } |
| 173 | - | ||
| 174 | int paragraphTokens = tokenEstimator.estimate(paragraph); | 167 | int paragraphTokens = tokenEstimator.estimate(paragraph); |
| 175 | - | ||
| 176 | // 如果单个段落就超过限制,需要按句子切分 | 168 | // 如果单个段落就超过限制,需要按句子切分 |
| 177 | if (paragraphTokens > availableTokens) { | 169 | if (paragraphTokens > availableTokens) { |
| 178 | // 先保存当前块 | 170 | // 先保存当前块 |
| @@ -181,22 +173,20 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -181,22 +173,20 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 181 | currentChunk = new StringBuilder(); | 173 | currentChunk = new StringBuilder(); |
| 182 | currentTokens = 0; | 174 | currentTokens = 0; |
| 183 | } | 175 | } |
| 184 | - | ||
| 185 | // 按句子切分长段落 | 176 | // 按句子切分长段落 |
| 186 | chunks.addAll(splitLongParagraph(paragraph, availableTokens)); | 177 | chunks.addAll(splitLongParagraph(paragraph, availableTokens)); |
| 187 | continue; | 178 | continue; |
| 188 | } | 179 | } |
| 189 | - | ||
| 190 | // 如果加上这个段落会超过限制 | 180 | // 如果加上这个段落会超过限制 |
| 191 | if (currentTokens + paragraphTokens > availableTokens && currentChunk.length() > 0) { | 181 | if (currentTokens + paragraphTokens > availableTokens && currentChunk.length() > 0) { |
| 192 | chunks.add(currentChunk.toString().trim()); | 182 | chunks.add(currentChunk.toString().trim()); |
| 193 | currentChunk = new StringBuilder(); | 183 | currentChunk = new StringBuilder(); |
| 194 | currentTokens = 0; | 184 | currentTokens = 0; |
| 195 | } | 185 | } |
| 196 | - | ||
| 197 | if (currentChunk.length() > 0) { | 186 | if (currentChunk.length() > 0) { |
| 198 | currentChunk.append("\n\n"); | 187 | currentChunk.append("\n\n"); |
| 199 | } | 188 | } |
| 189 | + // 添加段落 | ||
| 200 | currentChunk.append(paragraph); | 190 | currentChunk.append(paragraph); |
| 201 | currentTokens += paragraphTokens; | 191 | currentTokens += paragraphTokens; |
| 202 | } | 192 | } |
| @@ -205,27 +195,29 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -205,27 +195,29 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 205 | if (currentChunk.length() > 0) { | 195 | if (currentChunk.length() > 0) { |
| 206 | chunks.add(currentChunk.toString().trim()); | 196 | chunks.add(currentChunk.toString().trim()); |
| 207 | } | 197 | } |
| 208 | - | ||
| 209 | - return chunks.isEmpty() ? Collections.singletonList(answer) : chunks; | 198 | + return CollUtil.isEmpty(chunks) ? Collections.singletonList(answer) : chunks; |
| 210 | } | 199 | } |
| 211 | 200 | ||
| 212 | /** | 201 | /** |
| 213 | * 切分长段落(按句子) | 202 | * 切分长段落(按句子) |
| 203 | + * | ||
| 204 | + * @param paragraph 段落文本 | ||
| 205 | + * @param availableTokens 可用的 Token 数 | ||
| 206 | + * @return 切分后的文本片段列表 | ||
| 214 | */ | 207 | */ |
| 215 | private List<String> splitLongParagraph(String paragraph, int availableTokens) { | 208 | private List<String> splitLongParagraph(String paragraph, int availableTokens) { |
| 209 | + // 按句子切分 | ||
| 216 | List<String> chunks = new ArrayList<>(); | 210 | List<String> chunks = new ArrayList<>(); |
| 217 | String[] sentences = SENTENCE_PATTERN.split(paragraph); | 211 | String[] sentences = SENTENCE_PATTERN.split(paragraph); |
| 218 | 212 | ||
| 213 | + // 按句子累积切分 | ||
| 219 | StringBuilder currentChunk = new StringBuilder(); | 214 | StringBuilder currentChunk = new StringBuilder(); |
| 220 | int currentTokens = 0; | 215 | int currentTokens = 0; |
| 221 | - | ||
| 222 | for (String sentence : sentences) { | 216 | for (String sentence : sentences) { |
| 223 | if (StrUtil.isEmpty(sentence)) { | 217 | if (StrUtil.isEmpty(sentence)) { |
| 224 | continue; | 218 | continue; |
| 225 | } | 219 | } |
| 226 | - | ||
| 227 | int sentenceTokens = tokenEstimator.estimate(sentence); | 220 | int sentenceTokens = tokenEstimator.estimate(sentence); |
| 228 | - | ||
| 229 | // 如果单个句子就超过限制,强制切分 | 221 | // 如果单个句子就超过限制,强制切分 |
| 230 | if (sentenceTokens > availableTokens) { | 222 | if (sentenceTokens > availableTokens) { |
| 231 | if (currentChunk.length() > 0) { | 223 | if (currentChunk.length() > 0) { |
| @@ -236,47 +228,50 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -236,47 +228,50 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 236 | chunks.add(sentence.trim()); | 228 | chunks.add(sentence.trim()); |
| 237 | continue; | 229 | continue; |
| 238 | } | 230 | } |
| 239 | - | 231 | + // 如果加上这个句子会超过限制 |
| 240 | if (currentTokens + sentenceTokens > availableTokens && currentChunk.length() > 0) { | 232 | if (currentTokens + sentenceTokens > availableTokens && currentChunk.length() > 0) { |
| 241 | chunks.add(currentChunk.toString().trim()); | 233 | chunks.add(currentChunk.toString().trim()); |
| 242 | currentChunk = new StringBuilder(); | 234 | currentChunk = new StringBuilder(); |
| 243 | currentTokens = 0; | 235 | currentTokens = 0; |
| 244 | } | 236 | } |
| 245 | - | 237 | + // 添加句子 |
| 246 | currentChunk.append(sentence); | 238 | currentChunk.append(sentence); |
| 247 | currentTokens += sentenceTokens; | 239 | currentTokens += sentenceTokens; |
| 248 | } | 240 | } |
| 249 | 241 | ||
| 242 | + // 添加最后一块 | ||
| 250 | if (currentChunk.length() > 0) { | 243 | if (currentChunk.length() > 0) { |
| 251 | chunks.add(currentChunk.toString().trim()); | 244 | chunks.add(currentChunk.toString().trim()); |
| 252 | } | 245 | } |
| 253 | - | ||
| 254 | return chunks.isEmpty() ? Collections.singletonList(paragraph) : chunks; | 246 | return chunks.isEmpty() ? Collections.singletonList(paragraph) : chunks; |
| 255 | } | 247 | } |
| 256 | 248 | ||
| 257 | /** | 249 | /** |
| 258 | * 降级切分策略(当未识别到 QA 格式时) | 250 | * 降级切分策略(当未识别到 QA 格式时) |
| 251 | + * | ||
| 252 | + * @param content 文本内容 | ||
| 253 | + * @return 切分后的文本片段列表 | ||
| 259 | */ | 254 | */ |
| 260 | private List<String> fallbackSplit(String content) { | 255 | private List<String> fallbackSplit(String content) { |
| 256 | + // 按段落切分 | ||
| 261 | List<String> chunks = new ArrayList<>(); | 257 | List<String> chunks = new ArrayList<>(); |
| 262 | String[] paragraphs = content.split(PARAGRAPH_SEPARATOR); | 258 | String[] paragraphs = content.split(PARAGRAPH_SEPARATOR); |
| 263 | 259 | ||
| 260 | + // 按段落累积切分 | ||
| 264 | StringBuilder currentChunk = new StringBuilder(); | 261 | StringBuilder currentChunk = new StringBuilder(); |
| 265 | int currentTokens = 0; | 262 | int currentTokens = 0; |
| 266 | - | ||
| 267 | for (String paragraph : paragraphs) { | 263 | for (String paragraph : paragraphs) { |
| 268 | if (StrUtil.isEmpty(paragraph)) { | 264 | if (StrUtil.isEmpty(paragraph)) { |
| 269 | continue; | 265 | continue; |
| 270 | } | 266 | } |
| 271 | - | ||
| 272 | int paragraphTokens = tokenEstimator.estimate(paragraph); | 267 | int paragraphTokens = tokenEstimator.estimate(paragraph); |
| 273 | - | 268 | + // 如果加上这个段落会超过限制 |
| 274 | if (currentTokens + paragraphTokens > chunkSize && currentChunk.length() > 0) { | 269 | if (currentTokens + paragraphTokens > chunkSize && currentChunk.length() > 0) { |
| 275 | chunks.add(currentChunk.toString().trim()); | 270 | chunks.add(currentChunk.toString().trim()); |
| 276 | currentChunk = new StringBuilder(); | 271 | currentChunk = new StringBuilder(); |
| 277 | currentTokens = 0; | 272 | currentTokens = 0; |
| 278 | } | 273 | } |
| 279 | - | 274 | + // 添加段落 |
| 280 | if (currentChunk.length() > 0) { | 275 | if (currentChunk.length() > 0) { |
| 281 | currentChunk.append("\n\n"); | 276 | currentChunk.append("\n\n"); |
| 282 | } | 277 | } |
| @@ -284,33 +279,32 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -284,33 +279,32 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 284 | currentTokens += paragraphTokens; | 279 | currentTokens += paragraphTokens; |
| 285 | } | 280 | } |
| 286 | 281 | ||
| 282 | + // 添加最后一块 | ||
| 287 | if (currentChunk.length() > 0) { | 283 | if (currentChunk.length() > 0) { |
| 288 | chunks.add(currentChunk.toString().trim()); | 284 | chunks.add(currentChunk.toString().trim()); |
| 289 | } | 285 | } |
| 290 | - | ||
| 291 | return chunks.isEmpty() ? Collections.singletonList(content) : chunks; | 286 | return chunks.isEmpty() ? Collections.singletonList(content) : chunks; |
| 292 | } | 287 | } |
| 293 | 288 | ||
| 294 | /** | 289 | /** |
| 295 | * QA 对数据结构 | 290 | * QA 对数据结构 |
| 296 | */ | 291 | */ |
| 292 | + @AllArgsConstructor | ||
| 297 | private static class QaPair { | 293 | private static class QaPair { |
| 294 | + | ||
| 298 | String question; | 295 | String question; |
| 299 | String answer; | 296 | String answer; |
| 300 | String fullText; | 297 | String fullText; |
| 301 | 298 | ||
| 302 | - QaPair(String question, String answer, String fullText) { | ||
| 303 | - this.question = question; | ||
| 304 | - this.answer = answer; | ||
| 305 | - this.fullText = fullText; | ||
| 306 | - } | ||
| 307 | } | 299 | } |
| 308 | 300 | ||
| 309 | /** | 301 | /** |
| 310 | * Token 估算器接口 | 302 | * Token 估算器接口 |
| 311 | */ | 303 | */ |
| 312 | public interface TokenEstimator { | 304 | public interface TokenEstimator { |
| 305 | + | ||
| 313 | int estimate(String text); | 306 | int estimate(String text); |
| 307 | + | ||
| 314 | } | 308 | } |
| 315 | 309 | ||
| 316 | /** | 310 | /** |
| @@ -319,6 +313,7 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -319,6 +313,7 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 319 | * 英文:1 单词 ≈ 1.3 Token | 313 | * 英文:1 单词 ≈ 1.3 Token |
| 320 | */ | 314 | */ |
| 321 | private static class SimpleTokenEstimator implements TokenEstimator { | 315 | private static class SimpleTokenEstimator implements TokenEstimator { |
| 316 | + | ||
| 322 | @Override | 317 | @Override |
| 323 | public int estimate(String text) { | 318 | public int estimate(String text) { |
| 324 | if (StrUtil.isEmpty(text)) { | 319 | if (StrUtil.isEmpty(text)) { |
| @@ -327,14 +322,12 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -327,14 +322,12 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 327 | 322 | ||
| 328 | int chineseChars = 0; | 323 | int chineseChars = 0; |
| 329 | int englishWords = 0; | 324 | int englishWords = 0; |
| 330 | - | ||
| 331 | // 简单统计中英文 | 325 | // 简单统计中英文 |
| 332 | for (char c : text.toCharArray()) { | 326 | for (char c : text.toCharArray()) { |
| 333 | if (c >= 0x4E00 && c <= 0x9FA5) { | 327 | if (c >= 0x4E00 && c <= 0x9FA5) { |
| 334 | chineseChars++; | 328 | chineseChars++; |
| 335 | } | 329 | } |
| 336 | } | 330 | } |
| 337 | - | ||
| 338 | // 英文单词估算 | 331 | // 英文单词估算 |
| 339 | String[] words = text.split("\\s+"); | 332 | String[] words = text.split("\\s+"); |
| 340 | for (String word : words) { | 333 | for (String word : words) { |
| @@ -342,8 +335,8 @@ public class MarkdownQaSplitter extends TextSplitter { | @@ -342,8 +335,8 @@ public class MarkdownQaSplitter extends TextSplitter { | ||
| 342 | englishWords++; | 335 | englishWords++; |
| 343 | } | 336 | } |
| 344 | } | 337 | } |
| 345 | - | ||
| 346 | return chineseChars + (int) (englishWords * 1.3); | 338 | return chineseChars + (int) (englishWords * 1.3); |
| 347 | } | 339 | } |
| 348 | } | 340 | } |
| 341 | + | ||
| 349 | } | 342 | } |
yudao-module-ai/src/main/java/cn/iocoder/yudao/module/ai/service/knowledge/splitter/SemanticTextSplitter.java
| @@ -8,6 +8,7 @@ import java.util.ArrayList; | @@ -8,6 +8,7 @@ import java.util.ArrayList; | ||
| 8 | import java.util.Arrays; | 8 | import java.util.Arrays; |
| 9 | import java.util.Collections; | 9 | import java.util.Collections; |
| 10 | import java.util.List; | 10 | import java.util.List; |
| 11 | +import java.util.regex.Matcher; | ||
| 11 | import java.util.regex.Pattern; | 12 | import java.util.regex.Pattern; |
| 12 | 13 | ||
| 13 | /** | 14 | /** |
| @@ -72,12 +73,14 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -72,12 +73,14 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 72 | if (StrUtil.isEmpty(text)) { | 73 | if (StrUtil.isEmpty(text)) { |
| 73 | return Collections.emptyList(); | 74 | return Collections.emptyList(); |
| 74 | } | 75 | } |
| 75 | - | ||
| 76 | return splitTextRecursive(text); | 76 | return splitTextRecursive(text); |
| 77 | } | 77 | } |
| 78 | 78 | ||
| 79 | /** | 79 | /** |
| 80 | * 切分文本(递归策略) | 80 | * 切分文本(递归策略) |
| 81 | + * | ||
| 82 | + * @param text 待切分文本 | ||
| 83 | + * @return 切分后的文本块列表 | ||
| 81 | */ | 84 | */ |
| 82 | private List<String> splitTextRecursive(String text) { | 85 | private List<String> splitTextRecursive(String text) { |
| 83 | List<String> chunks = new ArrayList<>(); | 86 | List<String> chunks = new ArrayList<>(); |
| @@ -92,7 +95,6 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -92,7 +95,6 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 92 | // 尝试按不同分隔符切分 | 95 | // 尝试按不同分隔符切分 |
| 93 | List<String> splits = null; | 96 | List<String> splits = null; |
| 94 | String usedSeparator = null; | 97 | String usedSeparator = null; |
| 95 | - | ||
| 96 | for (String separator : PARAGRAPH_SEPARATORS) { | 98 | for (String separator : PARAGRAPH_SEPARATORS) { |
| 97 | if (text.contains(separator)) { | 99 | if (text.contains(separator)) { |
| 98 | splits = Arrays.asList(text.split(Pattern.quote(separator))); | 100 | splits = Arrays.asList(text.split(Pattern.quote(separator))); |
| @@ -109,18 +111,20 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -109,18 +111,20 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 109 | 111 | ||
| 110 | // 合并小片段 | 112 | // 合并小片段 |
| 111 | chunks = mergeSplits(splits, usedSeparator); | 113 | chunks = mergeSplits(splits, usedSeparator); |
| 112 | - | ||
| 113 | return chunks; | 114 | return chunks; |
| 114 | } | 115 | } |
| 115 | 116 | ||
| 116 | /** | 117 | /** |
| 117 | * 按句子切分 | 118 | * 按句子切分 |
| 119 | + * | ||
| 120 | + * @param text 待切分文本 | ||
| 121 | + * @return 句子列表 | ||
| 118 | */ | 122 | */ |
| 119 | private List<String> splitBySentences(String text) { | 123 | private List<String> splitBySentences(String text) { |
| 124 | + // 使用正则表达式匹配句子结束位置 | ||
| 120 | List<String> sentences = new ArrayList<>(); | 125 | List<String> sentences = new ArrayList<>(); |
| 121 | int lastEnd = 0; | 126 | int lastEnd = 0; |
| 122 | - | ||
| 123 | - java.util.regex.Matcher matcher = SENTENCE_END_PATTERN.matcher(text); | 127 | + Matcher matcher = SENTENCE_END_PATTERN.matcher(text); |
| 124 | while (matcher.find()) { | 128 | while (matcher.find()) { |
| 125 | String sentence = text.substring(lastEnd, matcher.end()).trim(); | 129 | String sentence = text.substring(lastEnd, matcher.end()).trim(); |
| 126 | if (StrUtil.isNotEmpty(sentence)) { | 130 | if (StrUtil.isNotEmpty(sentence)) { |
| @@ -136,12 +140,15 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -136,12 +140,15 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 136 | sentences.add(remaining); | 140 | sentences.add(remaining); |
| 137 | } | 141 | } |
| 138 | } | 142 | } |
| 139 | - | ||
| 140 | return sentences.isEmpty() ? Collections.singletonList(text) : sentences; | 143 | return sentences.isEmpty() ? Collections.singletonList(text) : sentences; |
| 141 | } | 144 | } |
| 142 | 145 | ||
| 143 | /** | 146 | /** |
| 144 | * 合并切分后的小片段 | 147 | * 合并切分后的小片段 |
| 148 | + * | ||
| 149 | + * @param splits 切分后的片段列表 | ||
| 150 | + * @param separator 片段间的分隔符 | ||
| 151 | + * @return 合并后的文本块列表 | ||
| 145 | */ | 152 | */ |
| 146 | private List<String> mergeSplits(List<String> splits, String separator) { | 153 | private List<String> mergeSplits(List<String> splits, String separator) { |
| 147 | List<String> chunks = new ArrayList<>(); | 154 | List<String> chunks = new ArrayList<>(); |
| @@ -152,9 +159,7 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -152,9 +159,7 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 152 | if (StrUtil.isEmpty(split)) { | 159 | if (StrUtil.isEmpty(split)) { |
| 153 | continue; | 160 | continue; |
| 154 | } | 161 | } |
| 155 | - | ||
| 156 | int splitTokens = tokenEstimator.estimate(split); | 162 | int splitTokens = tokenEstimator.estimate(split); |
| 157 | - | ||
| 158 | // 如果单个片段就超过限制,进一步递归切分 | 163 | // 如果单个片段就超过限制,进一步递归切分 |
| 159 | if (splitTokens > chunkSize) { | 164 | if (splitTokens > chunkSize) { |
| 160 | // 先保存当前累积的块 | 165 | // 先保存当前累积的块 |
| @@ -164,7 +169,6 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -164,7 +169,6 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 164 | currentChunks.clear(); | 169 | currentChunks.clear(); |
| 165 | currentLength = 0; | 170 | currentLength = 0; |
| 166 | } | 171 | } |
| 167 | - | ||
| 168 | // 递归切分大片段 | 172 | // 递归切分大片段 |
| 169 | if (!separator.isEmpty()) { | 173 | if (!separator.isEmpty()) { |
| 170 | // 如果是段落分隔符,尝试按句子切分 | 174 | // 如果是段落分隔符,尝试按句子切分 |
| @@ -175,10 +179,8 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -175,10 +179,8 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 175 | } | 179 | } |
| 176 | continue; | 180 | continue; |
| 177 | } | 181 | } |
| 178 | - | ||
| 179 | // 计算加上分隔符的 Token 数 | 182 | // 计算加上分隔符的 Token 数 |
| 180 | int separatorTokens = StrUtil.isEmpty(separator) ? 0 : tokenEstimator.estimate(separator); | 183 | int separatorTokens = StrUtil.isEmpty(separator) ? 0 : tokenEstimator.estimate(separator); |
| 181 | - | ||
| 182 | // 如果加上这个片段会超过限制 | 184 | // 如果加上这个片段会超过限制 |
| 183 | if (!currentChunks.isEmpty() && currentLength + splitTokens + separatorTokens > chunkSize) { | 185 | if (!currentChunks.isEmpty() && currentLength + splitTokens + separatorTokens > chunkSize) { |
| 184 | // 保存当前块 | 186 | // 保存当前块 |
| @@ -189,7 +191,7 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -189,7 +191,7 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 189 | currentChunks = getOverlappingChunks(currentChunks, separator); | 191 | currentChunks = getOverlappingChunks(currentChunks, separator); |
| 190 | currentLength = estimateTokens(currentChunks, separator); | 192 | currentLength = estimateTokens(currentChunks, separator); |
| 191 | } | 193 | } |
| 192 | - | 194 | + // 添加当前片段 |
| 193 | currentChunks.add(split); | 195 | currentChunks.add(split); |
| 194 | currentLength += splitTokens + separatorTokens; | 196 | currentLength += splitTokens + separatorTokens; |
| 195 | } | 197 | } |
| @@ -199,39 +201,43 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -199,39 +201,43 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 199 | String chunkText = String.join(separator, currentChunks); | 201 | String chunkText = String.join(separator, currentChunks); |
| 200 | chunks.add(chunkText.trim()); | 202 | chunks.add(chunkText.trim()); |
| 201 | } | 203 | } |
| 202 | - | ||
| 203 | return chunks; | 204 | return chunks; |
| 204 | } | 205 | } |
| 205 | 206 | ||
| 206 | /** | 207 | /** |
| 207 | * 获取重叠的片段(用于保持上下文) | 208 | * 获取重叠的片段(用于保持上下文) |
| 209 | + * | ||
| 210 | + * @param chunks 当前片段列表 | ||
| 211 | + * @param separator 片段间的分隔符 | ||
| 212 | + * @return 重叠的片段列表 | ||
| 208 | */ | 213 | */ |
| 209 | private List<String> getOverlappingChunks(List<String> chunks, String separator) { | 214 | private List<String> getOverlappingChunks(List<String> chunks, String separator) { |
| 210 | if (chunkOverlap == 0 || chunks.isEmpty()) { | 215 | if (chunkOverlap == 0 || chunks.isEmpty()) { |
| 211 | return new ArrayList<>(); | 216 | return new ArrayList<>(); |
| 212 | } | 217 | } |
| 213 | 218 | ||
| 219 | + // 从后往前取片段,直到达到重叠大小 | ||
| 214 | List<String> overlapping = new ArrayList<>(); | 220 | List<String> overlapping = new ArrayList<>(); |
| 215 | int tokens = 0; | 221 | int tokens = 0; |
| 216 | - | ||
| 217 | - // 从后往前取片段,直到达到重叠大小 | ||
| 218 | for (int i = chunks.size() - 1; i >= 0; i--) { | 222 | for (int i = chunks.size() - 1; i >= 0; i--) { |
| 219 | String chunk = chunks.get(i); | 223 | String chunk = chunks.get(i); |
| 220 | int chunkTokens = tokenEstimator.estimate(chunk); | 224 | int chunkTokens = tokenEstimator.estimate(chunk); |
| 221 | - | ||
| 222 | if (tokens + chunkTokens > chunkOverlap) { | 225 | if (tokens + chunkTokens > chunkOverlap) { |
| 223 | break; | 226 | break; |
| 224 | } | 227 | } |
| 225 | - | 228 | + // 添加到重叠列表前端 |
| 226 | overlapping.add(0, chunk); | 229 | overlapping.add(0, chunk); |
| 227 | tokens += chunkTokens + (StrUtil.isEmpty(separator) ? 0 : tokenEstimator.estimate(separator)); | 230 | tokens += chunkTokens + (StrUtil.isEmpty(separator) ? 0 : tokenEstimator.estimate(separator)); |
| 228 | } | 231 | } |
| 229 | - | ||
| 230 | return overlapping; | 232 | return overlapping; |
| 231 | } | 233 | } |
| 232 | 234 | ||
| 233 | /** | 235 | /** |
| 234 | * 估算片段列表的总 Token 数 | 236 | * 估算片段列表的总 Token 数 |
| 237 | + * | ||
| 238 | + * @param chunks 片段列表 | ||
| 239 | + * @param separator 片段间的分隔符 | ||
| 240 | + * @return 总 Token 数 | ||
| 235 | */ | 241 | */ |
| 236 | private int estimateTokens(List<String> chunks, String separator) { | 242 | private int estimateTokens(List<String> chunks, String separator) { |
| 237 | int total = 0; | 243 | int total = 0; |
| @@ -246,17 +252,18 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -246,17 +252,18 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 246 | 252 | ||
| 247 | /** | 253 | /** |
| 248 | * 强制切分长文本(当语义切分失败时) | 254 | * 强制切分长文本(当语义切分失败时) |
| 255 | + * | ||
| 256 | + * @param text 待切分文本 | ||
| 257 | + * @return 切分后的文本块列表 | ||
| 249 | */ | 258 | */ |
| 250 | private List<String> forceSplitLongText(String text) { | 259 | private List<String> forceSplitLongText(String text) { |
| 251 | List<String> chunks = new ArrayList<>(); | 260 | List<String> chunks = new ArrayList<>(); |
| 252 | int charsPerChunk = (int) (chunkSize * 0.8); // 保守估计 | 261 | int charsPerChunk = (int) (chunkSize * 0.8); // 保守估计 |
| 253 | - | ||
| 254 | for (int i = 0; i < text.length(); i += charsPerChunk) { | 262 | for (int i = 0; i < text.length(); i += charsPerChunk) { |
| 255 | int end = Math.min(i + charsPerChunk, text.length()); | 263 | int end = Math.min(i + charsPerChunk, text.length()); |
| 256 | String chunk = text.substring(i, end); | 264 | String chunk = text.substring(i, end); |
| 257 | chunks.add(chunk.trim()); | 265 | chunks.add(chunk.trim()); |
| 258 | } | 266 | } |
| 259 | - | ||
| 260 | log.warn("文本过长,已强制按字符切分,可能影响语义完整性"); | 267 | log.warn("文本过长,已强制按字符切分,可能影响语义完整性"); |
| 261 | return chunks; | 268 | return chunks; |
| 262 | } | 269 | } |
| @@ -265,6 +272,7 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -265,6 +272,7 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 265 | * 简单的 Token 估算器实现 | 272 | * 简单的 Token 估算器实现 |
| 266 | */ | 273 | */ |
| 267 | private static class SimpleTokenEstimator implements MarkdownQaSplitter.TokenEstimator { | 274 | private static class SimpleTokenEstimator implements MarkdownQaSplitter.TokenEstimator { |
| 275 | + | ||
| 268 | @Override | 276 | @Override |
| 269 | public int estimate(String text) { | 277 | public int estimate(String text) { |
| 270 | if (StrUtil.isEmpty(text)) { | 278 | if (StrUtil.isEmpty(text)) { |
| @@ -273,21 +281,21 @@ public class SemanticTextSplitter extends TextSplitter { | @@ -273,21 +281,21 @@ public class SemanticTextSplitter extends TextSplitter { | ||
| 273 | 281 | ||
| 274 | int chineseChars = 0; | 282 | int chineseChars = 0; |
| 275 | int englishWords = 0; | 283 | int englishWords = 0; |
| 276 | - | 284 | + // 简单统计中英文 |
| 277 | for (char c : text.toCharArray()) { | 285 | for (char c : text.toCharArray()) { |
| 278 | if (c >= 0x4E00 && c <= 0x9FA5) { | 286 | if (c >= 0x4E00 && c <= 0x9FA5) { |
| 279 | chineseChars++; | 287 | chineseChars++; |
| 280 | } | 288 | } |
| 281 | } | 289 | } |
| 282 | - | 290 | + // 英文单词估算 |
| 283 | String[] words = text.split("\\s+"); | 291 | String[] words = text.split("\\s+"); |
| 284 | for (String word : words) { | 292 | for (String word : words) { |
| 285 | if (word.matches(".*[a-zA-Z].*")) { | 293 | if (word.matches(".*[a-zA-Z].*")) { |
| 286 | englishWords++; | 294 | englishWords++; |
| 287 | } | 295 | } |
| 288 | } | 296 | } |
| 289 | - | ||
| 290 | return chineseChars + (int) (englishWords * 1.3); | 297 | return chineseChars + (int) (englishWords * 1.3); |
| 291 | } | 298 | } |
| 292 | } | 299 | } |
| 300 | + | ||
| 293 | } | 301 | } |