perf: text splitter (#4313)

* sync collection

* remove lock

* perf: text splitter

* update comment

docSite/content/zh-cn/docs/development/upgrading/492.md

@@ -30,6 +30,10 @@ weight: 799
 6. 工作流节点数组字符串类型，自动适配 string 输入。
 7. 工作流节点数组类型，自动进行 JSON parse 解析 string 输入。
 8. AI proxy 日志优化，去除重试失败的日志，仅保留最后一份错误日志。
+9. 分块算法小调整：  
+  * 跨处理符号之间连续性更强。  
+  * 代码块分割时，用 LLM 模型上下文作为分块大小，尽可能保证代码块完整性。
+  * 表格分割时，用 LLM 模型上下文作为分块大小，尽可能保证表格完整性。
 
 ## 🐛 修复
 
@@ -37,3 +41,4 @@ weight: 799
 2. 渠道测试时，如果配置了模型自定义请求地址，会走自定义请求地址，而不是渠道请求地址。
 3. 语音识别模型测试未启用的模型时，无法正常测试。
 4. 管理员配置系统插件时，如果插件包含其他系统应用，无法正常鉴权。
+5. 移除 TTS 自定义请求地址时，必须需要填 requestAuth 字段。

packages/global/common/string/textSplitter.ts

@@ -134,8 +134,11 @@ const commonSplit = (props: SplitProps): SplitResponse => {
     { reg: /^(####\s[^\n]+\n)/gm, maxLen: chunkSize },
     { reg: /^(#####\s[^\n]+\n)/gm, maxLen: chunkSize },
 
-    { reg: /([\n]([`~]))/g, maxLen: chunkSize }, // code block
+    { reg: /([\n](```[\s\S]*?```|~~~[\s\S]*?~~~))/g, maxLen: maxSize }, // code block
-    { reg: /([\n](?=\s*[0-9]+\.))/g, maxLen: chunkSize }, // 增大块，尽可能保证它是一个完整的段落。 (?![\*\-|>`0-9]): markdown special char
+    {
+      reg: /(\n\|(?:(?:[^\n|]+\|){1,})\n\|(?:[:\-\s]+\|){1,}\n(?:\|(?:[^\n|]+\|)*\n)*)/g,
+      maxLen: maxSize
+    }, // Table 尽可能保证完整性
     { reg: /(\n{2,})/g, maxLen: chunkSize },
     { reg: /([\n])/g, maxLen: chunkSize },
     // ------ There's no overlap on the top
@@ -150,7 +153,7 @@ const commonSplit = (props: SplitProps): SplitResponse => {
   const checkIsCustomStep = (step: number) => step < customRegLen;
   const checkIsMarkdownSplit = (step: number) =>
     step >= customRegLen && step <= markdownIndex + customRegLen;
-  +customReg.length;
+
   const checkForbidOverlap = (step: number) => step <= forbidOverlapIndex + customRegLen;
 
   // if use markdown title split, Separate record title
@@ -159,7 +162,8 @@ const commonSplit = (props: SplitProps): SplitResponse => {
       return [
         {
           text,
-          title: ''
+          title: '',
+          chunkMaxSize: chunkSize
         }
       ];
     }
@@ -167,7 +171,7 @@ const commonSplit = (props: SplitProps): SplitResponse => {
     const isCustomStep = checkIsCustomStep(step);
     const isMarkdownSplit = checkIsMarkdownSplit(step);
 
-    const { reg } = stepReges[step];
+    const { reg, maxLen } = stepReges[step];
 
     const replaceText = (() => {
       if (typeof reg === 'string') {
@@ -194,15 +198,19 @@ const commonSplit = (props: SplitProps): SplitResponse => {
         })()
       );
     })();
+
     const splitTexts = replaceText.split(splitMarker).filter((part) => part.trim());
 
     return splitTexts
       .map((text) => {
         const matchTitle = isMarkdownSplit ? text.match(reg)?.[0] || '' : '';
+        // 如果一个分块没有匹配到，则使用默认块大小，否则使用最大块大小
+        const chunkMaxSize = text.match(reg) === null ? chunkSize : maxLen;
 
         return {
           text: isMarkdownSplit ? text.replace(matchTitle, '') : text,
-          title: matchTitle
+          title: matchTitle,
+          chunkMaxSize
         };
       })
       .filter((item) => !!item.title || !!item.text?.trim());
@@ -252,9 +260,9 @@ const commonSplit = (props: SplitProps): SplitResponse => {
     const isCustomStep = checkIsCustomStep(step);
     const forbidConcat = isCustomStep; // forbid=true时候，lastText肯定为空
 
-    // oversize
+    // Over step
     if (step >= stepReges.length) {
-      if (text.length < chunkSize * 3) {
+      if (text.length < maxSize) {
         return [text];
       }
       // use slice-chunkSize to split text
@@ -268,19 +276,18 @@ const commonSplit = (props: SplitProps): SplitResponse => {
     // split text by special char
     const splitTexts = getSplitTexts({ text, step });
 
-    const maxLen = splitTexts.length > 1 ? stepReges[step].maxLen : chunkSize;
-    const minChunkLen = chunkSize * 0.7;
-
     const chunks: string[] = [];
     for (let i = 0; i < splitTexts.length; i++) {
       const item = splitTexts[i];
 
+      const maxLen = item.chunkMaxSize; // 当前块最大长度
+
       const lastTextLen = lastText.length;
       const currentText = item.text;
       const newText = lastText + currentText;
       const newTextLen = newText.length;
 
-      // Markdown 模式下，会强制向下拆分最小块，并再最后一个标题时候，给小块都补充上所有标题（包含父级标题）
+      // Markdown 模式下，会强制向下拆分最小块，并再最后一个标题深度，给小块都补充上所有标题（包含父级标题）
       if (isMarkdownStep) {
         // split new Text, split chunks must will greater 1 (small lastText)
         const innerChunks = splitTextRecursively({
@@ -290,11 +297,13 @@ const commonSplit = (props: SplitProps): SplitResponse => {
           parentTitle: parentTitle + item.title
         });
 
+        // 只有标题，没有内容。
         if (innerChunks.length === 0) {
           chunks.push(`${parentTitle}${item.title}`);
           continue;
         }
 
+        // 在合并最深级标题时，需要补充标题
         chunks.push(
           ...innerChunks.map(
             (chunk) =>
@@ -307,7 +316,16 @@ const commonSplit = (props: SplitProps): SplitResponse => {
 
       // newText is too large(now, The lastText must be smaller than chunkSize)
       if (newTextLen > maxLen) {
-        // lastText greater minChunkLen, direct push it to chunks, not add to next chunk. (large lastText)
+        const minChunkLen = maxLen * 0.8; // 当前块最小长度
+        const maxChunkLen = maxLen * 1.2; // 当前块最大长度
+
+        // 新文本没有非常大，直接认为它是一个新的块
+        if (newTextLen < maxChunkLen) {
+          chunks.push(newText);
+          lastText = getOneTextOverlapText({ text: newText, step }); // next chunk will start with overlayText
+          continue;
+        }
+        // 上一个文本块已经挺大的，单独做一个块
         if (lastTextLen > minChunkLen) {
           chunks.push(lastText);
 
@@ -317,13 +335,13 @@ const commonSplit = (props: SplitProps): SplitResponse => {
           continue;
         }
 
-        // 说明是新的文本块比较大，需要进一步拆分
+        // 说明是当前文本比较大，需要进一步拆分
 
-        // split new Text, split chunks must will greater 1 (small lastText)
+        // 把新的文本块进行一个拆分，并追加到 latestText 中
         const innerChunks = splitTextRecursively({
-          text: newText,
+          text: currentText,
           step: step + 1,
-          lastText: '',
+          lastText,
           parentTitle: parentTitle + item.title
         });
         const lastChunk = innerChunks[innerChunks.length - 1];
@@ -351,11 +369,11 @@ const commonSplit = (props: SplitProps): SplitResponse => {
 
       // Not overlap
       if (forbidConcat) {
-        chunks.push(item.text);
+        chunks.push(currentText);
         continue;
       }
 
-      lastText += item.text;
+      lastText = newText;
     }
 
     /* If the last chunk is independent, it needs to be push chunks. */

packages/service/core/ai/audio/speech.ts

@@ -30,11 +30,11 @@ export async function text2Speech({
       response_format: 'mp3',
       speed
     },
-    modelData.requestUrl && modelData.requestAuth
+    modelData.requestUrl
       ? {
           path: modelData.requestUrl,
           headers: {
-            Authorization: `Bearer ${modelData.requestAuth}`
+            ...(modelData.requestAuth ? { Authorization: `Bearer ${modelData.requestAuth}` } : {})
           }
         }
       : {}

packages/service/core/ai/utils.ts

@@ -65,6 +65,7 @@ export const llmCompletionsBodyFormat = <T extends CompletionsBodyType>(
 
   const requestBody: T = {
     ...body,
+    model: modelData.model,
     temperature:
       typeof body.temperature === 'number'
         ? computedTemperature({

projects/app/src/pageComponents/account/team/OrgManage/index.tsx

@@ -196,7 +196,7 @@ function OrgTable({ Tabs }: { Tabs: React.ReactNode }) {
         isLoading={isLoadingOrgs}
       >
         <Box mb={3}>
-          <Path paths={paths} rootName={userInfo?.team?.teamName} onClick={setPath} />
+          <Path paths={paths} rootName={userInfo?.team?.teamName} />
         </Box>
         <Flex flex={'1 0 0'} h={0} w={'100%'} gap={'4'}>
           <MemberScrollData flex="1">
@@ -420,7 +420,7 @@ function OrgTable({ Tabs }: { Tabs: React.ReactNode }) {
                   <ActionButton
                     icon="common/administrator"
                     text={t('account_team:manage_member')}
-                    onClick={() => setManageMemberOrg(currentOrg ?? rootOrg)}
+                    onClick={() => setManageMemberOrg(currentOrg)}
                   />
                   {currentOrg && currentOrg?.path !== '' && (
                     <>

projects/app/src/pages/api/core/dataset/file/getPreviewChunks.ts

@@ -94,7 +94,7 @@ async function handler(
     per: WritePermissionVal
   });
 
-  if (fileAuthRes && (String(fileAuthRes.tmbId) !== String(tmbId) || !fileAuthRes.isRoot)) {
+  if (fileAuthRes && String(fileAuthRes.tmbId) !== String(tmbId) && !fileAuthRes.isRoot) {
     return Promise.reject(CommonErrEnum.unAuthFile);
   }
 

projects/app/src/service/core/dataset/data/controller.ts

@@ -15,18 +15,19 @@ import { MongoDatasetDataText } from '@fastgpt/service/core/dataset/data/dataTex
 import { DatasetDataIndexTypeEnum } from '@fastgpt/global/core/dataset/data/constants';
 import { splitText2Chunks } from '@fastgpt/global/common/string/textSplitter';
 import { countPromptTokens } from '@fastgpt/service/common/string/tiktoken';
-import { getLLMMaxChunkSize } from '@fastgpt/global/core/dataset/training/utils';
 
 const formatIndexes = async ({
   indexes,
   q,
   a = '',
-  indexSize
+  indexSize,
+  maxIndexSize
 }: {
   indexes?: (Omit<DatasetDataIndexItemType, 'dataId'> & { dataId?: string })[];
   q: string;
   a?: string;
   indexSize: number;
+  maxIndexSize: number;
 }): Promise<
   {
     type: `${DatasetDataIndexTypeEnum}`;
@@ -46,9 +47,12 @@ const formatIndexes = async ({
   }) => {
     const qChunks = splitText2Chunks({
       text: q,
-      chunkSize: indexSize
+      chunkSize: indexSize,
+      maxSize: maxIndexSize
     }).chunks;
-    const aChunks = a ? splitText2Chunks({ text: a, chunkSize: indexSize }).chunks : [];
+    const aChunks = a
+      ? splitText2Chunks({ text: a, chunkSize: indexSize, maxSize: maxIndexSize }).chunks
+      : [];
 
     return [
       ...qChunks.map((text) => ({
@@ -100,7 +104,11 @@ const formatIndexes = async ({
         // If oversize tokens, split it
         const tokens = await countPromptTokens(item.text);
         if (tokens > indexSize) {
-          const splitText = splitText2Chunks({ text: item.text, chunkSize: 512 }).chunks;
+          const splitText = splitText2Chunks({
+            text: item.text,
+            chunkSize: 512,
+            maxSize: maxIndexSize
+          }).chunks;
           return splitText.map((text) => ({
             text,
             type: item.type
@@ -151,7 +159,8 @@ export async function insertData2Dataset({
     indexes,
     q,
     a,
-    indexSize
+    indexSize,
+    maxIndexSize: embModel.maxToken
   });
 
   // insert to vector store
@@ -236,7 +245,13 @@ export async function updateData2Dataset({
   if (!mongoData) return Promise.reject('core.dataset.error.Data not found');
 
   // 2. Compute indexes
-  const formatIndexesResult = await formatIndexes({ indexes, q, a, indexSize });
+  const formatIndexesResult = await formatIndexes({
+    indexes,
+    q,
+    a,
+    indexSize,
+    maxIndexSize: getEmbeddingModel(model).maxToken
+  });
 
   // 3. Patch indexes, create, update, delete
   const patchResult: PatchIndexesProps[] = [];