perf: chunk trigger and paragraph split (#4893)

* perf: chunk trigger and paragraph split

* update max size computed

* perf: i18n

* remove table

Author: c121914yu

docSite/content/zh-cn/docs/development/upgrading/4910.md

@@ -11,9 +11,11 @@ weight: 790
 ## 🚀 新增内容
 
 1. 支持 PG 设置`systemEnv.hnswMaxScanTuples`参数，提高迭代搜索的数据总量。
-2. 工作流调整为单向接入和接出，支持快速的添加下一步节点。
-3. 开放飞书和语雀知识库到开源版。
-4. gemini 和 claude 最新模型预设。
+2. 知识库预处理参数增加 “分块条件”，可控制某些情况下不进行分块处理。
+3. 知识库预处理参数增加 “段落优先” 模式，可控制最大段落深度。原“长度优先”模式，不再内嵌段落优先逻辑。
+4. 工作流调整为单向接入和接出，支持快速的添加下一步节点。
+5. 开放飞书和语雀知识库到开源版。
+6. gemini 和 claude 最新模型预设。
 
 ## ⚙️ 优化
 
@@ -31,4 +33,5 @@ weight: 790
 3. 工具调用模式，未保存思考输出。
 4. 知识库 indexSize 参数未生效。
 5. 工作流嵌套 2 层后，获取预览引用、上下文不正确。
-6. xlsx 转成 Markdown 时候，前面会多出一个空格。
+6. xlsx 转成 Markdown 时候，前面会多出一个空格。
+7. 读取 Markdown 文件时，Base64 图片未进行额外抓换保存。

packages/global/common/string/textSplitter.ts

@@ -7,6 +7,10 @@ export const CUSTOM_SPLIT_SIGN = '-----CUSTOM_SPLIT_SIGN-----';
 type SplitProps = {
   text: string;
   chunkSize: number;
+
+  paragraphChunkDeep?: number; // Paragraph deep
+  paragraphChunkMinSize?: number; // Paragraph min size, if too small, it will merge
+
   maxSize?: number;
   overlapRatio?: number;
   customReg?: string[];
@@ -108,6 +112,8 @@ const commonSplit = (props: SplitProps): SplitResponse => {
   let {
     text = '',
     chunkSize,
+    paragraphChunkDeep = 5,
+    paragraphChunkMinSize = 100,
     maxSize = defaultMaxChunkSize,
     overlapRatio = 0.15,
     customReg = []
@@ -123,7 +129,7 @@ const commonSplit = (props: SplitProps): SplitResponse => {
   text = text.replace(/(```[\s\S]*?```|~~~[\s\S]*?~~~)/g, function (match) {
     return match.replace(/\n/g, codeBlockMarker);
   });
-  // 2. 表格处理 - 单独提取表格出来，进行表头合并
+  // 2. Markdown 表格处理 - 单独提取表格出来，进行表头合并
   const tableReg =
     /(\n\|(?:(?:[^\n|]+\|){1,})\n\|(?:[:\-\s]+\|){1,}\n(?:\|(?:[^\n|]+\|)*\n?)*)(?:\n|$)/g;
   const tableDataList = text.match(tableReg);
@@ -143,25 +149,40 @@ const commonSplit = (props: SplitProps): SplitResponse => {
   text = text.replace(/(\r?\n|\r){3,}/g, '\n\n\n');
 
   // The larger maxLen is, the next sentence is less likely to trigger splitting
-  const markdownIndex = 4;
-  const forbidOverlapIndex = 8;
+  const customRegLen = customReg.length;
+  const markdownIndex = paragraphChunkDeep - 1;
+  const forbidOverlapIndex = customRegLen + markdownIndex + 4;
+
+  const markdownHeaderRules = ((deep?: number): { reg: RegExp; maxLen: number }[] => {
+    if (!deep || deep === 0) return [];
+
+    const maxDeep = Math.min(deep, 8); // Maximum 8 levels
+    const rules: { reg: RegExp; maxLen: number }[] = [];
+
+    for (let i = 1; i <= maxDeep; i++) {
+      const hashSymbols = '#'.repeat(i);
+      rules.push({
+        reg: new RegExp(`^(${hashSymbols}\\s[^\\n]+\\n)`, 'gm'),
+        maxLen: chunkSize
+      });
+    }
+
+    return rules;
+  })(paragraphChunkDeep);
 
   const stepReges: { reg: RegExp | string; maxLen: number }[] = [
     ...customReg.map((text) => ({
       reg: text.replaceAll('\\n', '\n'),
       maxLen: chunkSize
     })),
-    { reg: /^(#\s[^\n]+\n)/gm, maxLen: chunkSize },
-    { reg: /^(##\s[^\n]+\n)/gm, maxLen: chunkSize },
-    { reg: /^(###\s[^\n]+\n)/gm, maxLen: chunkSize },
-    { reg: /^(####\s[^\n]+\n)/gm, maxLen: chunkSize },
-    { reg: /^(#####\s[^\n]+\n)/gm, maxLen: chunkSize },
+    ...markdownHeaderRules,
 
     { reg: /([\n](```[\s\S]*?```|~~~[\s\S]*?~~~))/g, maxLen: maxSize }, // code block
+    // HTML Table tag 尽可能保障完整
     {
       reg: /(\n\|(?:(?:[^\n|]+\|){1,})\n\|(?:[:\-\s]+\|){1,}\n(?:\|(?:[^\n|]+\|)*\n)*)/g,
-      maxLen: Math.min(chunkSize * 1.5, maxSize)
-    }, // Table 尽可能保证完整性
+      maxLen: chunkSize
+    }, // Markdown 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
@@ -172,12 +193,10 @@ const commonSplit = (props: SplitProps): SplitResponse => {
     { reg: /([，]|,\s)/g, maxLen: chunkSize }
   ];
 
-  const customRegLen = customReg.length;
   const checkIsCustomStep = (step: number) => step < customRegLen;
   const checkIsMarkdownSplit = (step: number) =>
     step >= customRegLen && step <= markdownIndex + customRegLen;
-
-  const checkForbidOverlap = (step: number) => step <= forbidOverlapIndex + customRegLen;
+  const checkForbidOverlap = (step: number) => step <= forbidOverlapIndex;
 
   // if use markdown title split, Separate record title
   const getSplitTexts = ({ text, step }: { text: string; step: number }) => {
@@ -301,6 +320,7 @@ const commonSplit = (props: SplitProps): SplitResponse => {
     const splitTexts = getSplitTexts({ text, step });
 
     const chunks: string[] = [];
+
     for (let i = 0; i < splitTexts.length; i++) {
       const item = splitTexts[i];
 
@@ -443,7 +463,6 @@ const commonSplit = (props: SplitProps): SplitResponse => {
  */
 export const splitText2Chunks = (props: SplitProps): SplitResponse => {
   let { text = '' } = props;
-  const start = Date.now();
   const splitWithCustomSign = text.split(CUSTOM_SPLIT_SIGN);
 
   const splitResult = splitWithCustomSign.map((item) => {

packages/global/core/dataset/training/utils.ts

@@ -120,7 +120,6 @@ export const computeChunkSize = (params: {
 
   return Math.min(params.chunkSize ?? chunkAutoChunkSize, getLLMMaxChunkSize(params.llmModel));
 };
-
 export const computeChunkSplitter = (params: {
   chunkSettingMode?: ChunkSettingModeEnum;
   chunkSplitMode?: DataChunkSplitModeEnum;
@@ -129,8 +128,21 @@ export const computeChunkSplitter = (params: {
   if (params.chunkSettingMode === ChunkSettingModeEnum.auto) {
     return undefined;
   }
-  if (params.chunkSplitMode === DataChunkSplitModeEnum.size) {
+  if (params.chunkSplitMode !== DataChunkSplitModeEnum.char) {
     return undefined;
   }
   return params.chunkSplitter;
 };
+export const computeParagraphChunkDeep = (params: {
+  chunkSettingMode?: ChunkSettingModeEnum;
+  chunkSplitMode?: DataChunkSplitModeEnum;
+  paragraphChunkDeep?: number;
+}) => {
+  if (params.chunkSettingMode === ChunkSettingModeEnum.auto) {
+    return 5;
+  }
+  if (params.chunkSplitMode === DataChunkSplitModeEnum.paragraph) {
+    return params.paragraphChunkDeep;
+  }
+  return 0;
+};

packages/global/core/dataset/type.d.ts

@@ -9,7 +9,8 @@ import type {
   DatasetTypeEnum,
   SearchScoreTypeEnum,
   TrainingModeEnum,
-  ChunkSettingModeEnum
+  ChunkSettingModeEnum,
+  ChunkTriggerConfigTypeEnum
 } from './constants';
 import type { DatasetPermission } from '../../support/permission/dataset/controller';
 import type { APIFileServer, FeishuServer, YuqueServer } from './apiDataset';
@@ -37,11 +38,10 @@ export type ChunkSettingsType = {
   paragraphChunkAIMode?: ParagraphChunkAIModeEnum;
   paragraphChunkDeep?: number; // Paragraph deep
   paragraphChunkMinSize?: number; // Paragraph min size, if too small, it will merge
-  paragraphChunkMaxSize?: number; // Paragraph max size, if too large, it will split
   // Size split
-  chunkSize?: number;
+  chunkSize?: number; // chunk/qa chunk size, Paragraph max chunk size.
   // Char split
-  chunkSplitter?: string;
+  chunkSplitter?: string; // chunk/qa chunk splitter
   indexSize?: number;
 
   qaPrompt?: string;

packages/service/core/dataset/collection/controller.ts

@@ -34,6 +34,7 @@ import { getTrainingModeByCollection } from './utils';
 import {
   computeChunkSize,
   computeChunkSplitter,
+  computeParagraphChunkDeep,
   getLLMMaxChunkSize
 } from '@fastgpt/global/core/dataset/training/utils';
 import { DatasetDataIndexTypeEnum } from '@fastgpt/global/core/dataset/data/constants';
@@ -74,6 +75,8 @@ export const createCollectionAndInsertData = async ({
     llmModel: getLLMModel(dataset.agentModel)
   });
   const chunkSplitter = computeChunkSplitter(createCollectionParams);
+  const paragraphChunkDeep = computeParagraphChunkDeep(createCollectionParams);
+
   if (trainingType === DatasetCollectionDataProcessModeEnum.qa) {
     delete createCollectionParams.chunkTriggerType;
     delete createCollectionParams.chunkTriggerMinSize;
@@ -87,7 +90,11 @@ export const createCollectionAndInsertData = async ({
   // 1. split chunks
   const chunks = rawText2Chunks({
     rawText,
+    chunkTriggerType: createCollectionParams.chunkTriggerType,
+    chunkTriggerMinSize: createCollectionParams.chunkTriggerMinSize,
     chunkSize,
+    paragraphChunkDeep,
+    paragraphChunkMinSize: createCollectionParams.paragraphChunkMinSize,
     maxSize: getLLMMaxChunkSize(getLLMModel(dataset.agentModel)),
     overlapRatio: trainingType === DatasetCollectionDataProcessModeEnum.chunk ? 0.2 : 0,
     customReg: chunkSplitter ? [chunkSplitter] : [],
@@ -112,6 +119,7 @@ export const createCollectionAndInsertData = async ({
     const { _id: collectionId } = await createOneCollection({
       ...createCollectionParams,
       trainingType,
+      paragraphChunkDeep,
       chunkSize,
       chunkSplitter,
 
@@ -212,88 +220,38 @@ export type CreateOneCollectionParams = CreateDatasetCollectionParams & {
   tmbId: string;
   session?: ClientSession;
 };
-export async function createOneCollection({
-  teamId,
-  tmbId,
-  name,
-  parentId,
-  datasetId,
-  type,
-
-  createTime,
-  updateTime,
-
-  hashRawText,
-  rawTextLength,
-  metadata = {},
-  tags,
-
-  nextSyncTime,
-
-  fileId,
-  rawLink,
-  externalFileId,
-  externalFileUrl,
-  apiFileId,
-
-  // Parse settings
-  customPdfParse,
-  imageIndex,
-  autoIndexes,
-
-  // Chunk settings
-  trainingType,
-  chunkSettingMode,
-  chunkSplitMode,
-  chunkSize,
-  indexSize,
-  chunkSplitter,
-  qaPrompt,
-
-  session
-}: CreateOneCollectionParams) {
+export async function createOneCollection({ session, ...props }: CreateOneCollectionParams) {
+  const {
+    teamId,
+    parentId,
+    datasetId,
+    tags,
+
+    fileId,
+    rawLink,
+    externalFileId,
+    externalFileUrl,
+    apiFileId
+  } = props;
   // Create collection tags
   const collectionTags = await createOrGetCollectionTags({ tags, teamId, datasetId, session });
 
   // Create collection
   const [collection] = await MongoDatasetCollection.create(
     [
       {
+        ...props,
         teamId,
-        tmbId,
         parentId: parentId || null,
         datasetId,
-        name,
-        type,
 
-        rawTextLength,
-        hashRawText,
         tags: collectionTags,
-        metadata,
-
-        createTime,
-        updateTime,
-        nextSyncTime,
 
         ...(fileId ? { fileId } : {}),
         ...(rawLink ? { rawLink } : {}),
         ...(externalFileId ? { externalFileId } : {}),
         ...(externalFileUrl ? { externalFileUrl } : {}),
-        ...(apiFileId ? { apiFileId } : {}),
-
-        // Parse settings
-        customPdfParse,
-        imageIndex,
-        autoIndexes,
-
-        // Chunk settings
-        trainingType,
-        chunkSettingMode,
-        chunkSplitMode,
-        chunkSize,
-        indexSize,
-        chunkSplitter,
-        qaPrompt
+        ...(apiFileId ? { apiFileId } : {})
       }
     ],
     { session, ordered: true }

packages/service/core/dataset/read.ts

@@ -1,5 +1,8 @@
 import { BucketNameEnum } from '@fastgpt/global/common/file/constants';
-import { DatasetSourceReadTypeEnum } from '@fastgpt/global/core/dataset/constants';
+import {
+  ChunkTriggerConfigTypeEnum,
+  DatasetSourceReadTypeEnum
+} from '@fastgpt/global/core/dataset/constants';
 import { readFileContentFromMongo } from '../../common/file/gridfs/controller';
 import { urlsFetch } from '../../common/string/cheerio';
 import { type TextSplitProps, splitText2Chunks } from '@fastgpt/global/common/string/textSplitter';
@@ -179,11 +182,17 @@ export const readApiServerFileContent = async ({
 
 export const rawText2Chunks = ({
   rawText,
+  chunkTriggerType = ChunkTriggerConfigTypeEnum.minSize,
+  chunkTriggerMinSize = 1000,
   backupParse,
   chunkSize = 512,
   ...splitProps
 }: {
   rawText: string;
+
+  chunkTriggerType?: ChunkTriggerConfigTypeEnum;
+  chunkTriggerMinSize?: number; // maxSize from agent model, not store
+
   backupParse?: boolean;
   tableParse?: boolean;
 } & TextSplitProps): {
@@ -209,6 +218,28 @@ export const rawText2Chunks = ({
     };
   };
 
+  // Chunk condition
+  // 1. 选择最大值条件，只有超过了最大值(默认为模型的最大值*0.7），才会触发分块
+  if (chunkTriggerType === ChunkTriggerConfigTypeEnum.maxSize) {
+    const textLength = rawText.trim().length;
+    const maxSize = splitProps.maxSize ? splitProps.maxSize * 0.7 : 16000;
+    if (textLength < maxSize) {
+      return [
+        {
+          q: rawText,
+          a: ''
+        }
+      ];
+    }
+  }
+  // 2. 选择最小值条件，只有超过最小值(手动决定)才会触发分块
+  if (chunkTriggerType !== ChunkTriggerConfigTypeEnum.forceChunk) {
+    const textLength = rawText.trim().length;
+    if (textLength < chunkTriggerMinSize) {
+      return [{ q: rawText, a: '' }];
+    }
+  }
+
   if (backupParse) {
     return parseDatasetBackup2Chunks(rawText).chunks;
   }

packages/service/core/dataset/schema.ts

@@ -47,7 +47,6 @@ export const ChunkSettings = {
   },
   paragraphChunkDeep: Number,
   paragraphChunkMinSize: Number,
-  paragraphChunkMaxSize: Number,
   chunkSize: Number,
   chunkSplitter: String,