V4.6.6-1 (#656)

packages/global/common/file/read/index.ts

@@ -0,0 +1,62 @@
+/* read file to txt */
+import * as pdfjsLib from 'pdfjs-dist';
+
+export const readPdfFile = async ({ pdf }: { pdf: string | URL | ArrayBuffer }) => {
+  pdfjsLib.GlobalWorkerOptions.workerSrc = '/js/pdf.worker.js';
+
+  type TokenType = {
+    str: string;
+    dir: string;
+    width: number;
+    height: number;
+    transform: number[];
+    fontName: string;
+    hasEOL: boolean;
+  };
+
+  const readPDFPage = async (doc: any, pageNo: number) => {
+    const page = await doc.getPage(pageNo);
+    const tokenizedText = await page.getTextContent();
+
+    const viewport = page.getViewport({ scale: 1 });
+    const pageHeight = viewport.height;
+    const headerThreshold = pageHeight * 0.07; // 假设页头在页面顶部5%的区域内
+    const footerThreshold = pageHeight * 0.93; // 假设页脚在页面底部5%的区域内
+
+    const pageTexts: TokenType[] = tokenizedText.items.filter((token: TokenType) => {
+      return (
+        !token.transform ||
+        (token.transform[5] > headerThreshold && token.transform[5] < footerThreshold)
+      );
+    });
+
+    // concat empty string 'hasEOL'
+    for (let i = 0; i < pageTexts.length; i++) {
+      const item = pageTexts[i];
+      if (item.str === '' && pageTexts[i - 1]) {
+        pageTexts[i - 1].hasEOL = item.hasEOL;
+        pageTexts.splice(i, 1);
+        i--;
+      }
+    }
+
+    page.cleanup();
+
+    return pageTexts
+      .map((token) => {
+        const paragraphEnd = token.hasEOL && /([。？！.?!\n\r]|(\r\n))$/.test(token.str);
+
+        return paragraphEnd ? `${token.str}\n` : token.str;
+      })
+      .join('');
+  };
+
+  const doc = await pdfjsLib.getDocument(pdf).promise;
+  const pageTextPromises = [];
+  for (let pageNo = 1; pageNo <= doc.numPages; pageNo++) {
+    pageTextPromises.push(readPDFPage(doc, pageNo));
+  }
+  const pageTexts = await Promise.all(pageTextPromises);
+
+  return pageTexts.join('');
+};

packages/global/common/string/markdown.ts

@@ -34,3 +34,41 @@ export const simpleMarkdownText = (rawText: string) => {
 
   return rawText.trim();
 };
+
+/**
+ * format markdown
+ * 1. upload base64
+ * 2. replace \
+ */
+export const uploadMarkdownBase64 = async ({
+  rawText,
+  uploadImgController
+}: {
+  rawText: string;
+  uploadImgController: (base64: string) => Promise<string>;
+}) => {
+  // match base64, upload and replace it
+  const base64Regex = /data:image\/.*;base64,([^\)]+)/g;
+  const base64Arr = rawText.match(base64Regex) || [];
+  // upload base64 and replace it
+  await Promise.all(
+    base64Arr.map(async (base64Img) => {
+      try {
+        const str = await uploadImgController(base64Img);
+
+        rawText = rawText.replace(base64Img, str);
+      } catch (error) {
+        rawText = rawText.replace(base64Img, '');
+        rawText = rawText.replace(/!\[.*\]\(\)/g, '');
+      }
+    })
+  );
+
+  // Remove white space on both sides of the picture
+  const trimReg = /(!\[.*\]\(.*\))\s*/g;
+  if (trimReg.test(rawText)) {
+    rawText = rawText.replace(trimReg, '$1');
+  }
+
+  return simpleMarkdownText(rawText);
+};

packages/global/common/string/textSplitter.ts

@@ -31,7 +31,7 @@ export const splitText2Chunks = (props: {
 
   // The larger maxLen is, the next sentence is less likely to trigger splitting
   const stepReges: { reg: RegExp; maxLen: number }[] = [
-    ...customReg.map((text) => ({ reg: new RegExp(`([${text}])`, 'g'), maxLen: chunkLen * 1.4 })),
+    ...customReg.map((text) => ({ reg: new RegExp(`(${text})`, 'g'), maxLen: chunkLen * 1.4 })),
     { reg: /^(#\s[^\n]+)\n/gm, maxLen: chunkLen * 1.2 },
     { reg: /^(##\s[^\n]+)\n/gm, maxLen: chunkLen * 1.2 },
     { reg: /^(###\s[^\n]+)\n/gm, maxLen: chunkLen * 1.2 },
@@ -64,13 +64,22 @@ export const splitText2Chunks = (props: {
         }
       ];
     }
+
+    const isCustomSteep = checkIsCustomStep(step);
     const isMarkdownSplit = checkIsMarkdownSplit(step);
     const independentChunk = checkIndependentChunk(step);
 
     const { reg } = stepReges[step];
 
     const splitTexts = text
-      .replace(reg, independentChunk ? `${splitMarker}$1` : `$1${splitMarker}`)
+      .replace(
+        reg,
+        (() => {
+          if (isCustomSteep) return splitMarker;
+          if (independentChunk) return `${splitMarker}$1`;
+          return `$1${splitMarker}`;
+        })()
+      )
       .split(`${splitMarker}`)
       .filter((part) => part.trim());
 
@@ -128,11 +137,6 @@ export const splitText2Chunks = (props: {
     const independentChunk = checkIndependentChunk(step);
     const isCustomStep = checkIsCustomStep(step);
 
-    // mini text
-    if (text.length <= chunkLen) {
-      return [text];
-    }
-
     // oversize
     if (step >= stepReges.length) {
       if (text.length < chunkLen * 3) {
@@ -221,6 +225,8 @@ export const splitText2Chunks = (props: {
       } else {
         chunks.push(`${mdTitle}${lastText}`);
       }
+    } else if (lastText && chunks.length === 0) {
+      chunks.push(lastText);
     }
 
     return chunks;

packages/global/common/system/types/index.d.ts

@@ -1,4 +1,29 @@
-export type FeConfigsType = {
+import type {
+  ChatModelItemType,
+  FunctionModelItemType,
+  LLMModelItemType,
+  VectorModelItemType,
+  AudioSpeechModels,
+  WhisperModelType,
+  ReRankModelItemType
+} from '../../../core/ai/model.d';
+
+/* fastgpt main */
+export type FastGPTConfigFileType = {
+  feConfigs: FastGPTFeConfigsType;
+  systemEnv: SystemEnvType;
+  chatModels: ChatModelItemType[];
+  qaModels: LLMModelItemType[];
+  cqModels: FunctionModelItemType[];
+  extractModels: FunctionModelItemType[];
+  qgModels: LLMModelItemType[];
+  vectorModels: VectorModelItemType[];
+  reRankModels: ReRankModelItemType[];
+  audioSpeechModels: AudioSpeechModelType[];
+  whisperModel: WhisperModelType;
+};
+
+export type FastGPTFeConfigsType = {
   show_emptyChat?: boolean;
   show_register?: boolean;
   show_appStore?: boolean;
@@ -34,6 +59,6 @@ export type SystemEnvType = {
 };
 
 declare global {
-  var feConfigs: FeConfigsType;
+  var feConfigs: FastGPTFeConfigsType;
   var systemEnv: SystemEnvType;
 }

packages/global/core/ai/model.d.ts

@@ -24,6 +24,7 @@ export type VectorModelItemType = {
   defaultToken: number;
   price: number;
   maxToken: number;
+  weight: number;
 };
 
 export type ReRankModelItemType = {

packages/global/core/ai/model.ts

@@ -16,6 +16,7 @@ export const defaultVectorModels: VectorModelItemType[] = [
     name: 'Embedding-2',
     price: 0,
     defaultToken: 500,
-    maxToken: 3000
+    maxToken: 3000,
+    weight: 100
   }
 ];

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

@@ -89,6 +89,7 @@ export type DatasetTrainingSchemaType = {
   q: string;
   a: string;
   chunkIndex: number;
+  weight: number;
   indexes: Omit<DatasetDataIndexItemType, 'dataId'>[];
 };
 

packages/global/core/module/template/system/contextExtract.ts

@@ -36,10 +36,11 @@ export const ContextExtractModule: FlowModuleTemplateType = {
       type: FlowNodeInputTypeEnum.textarea,
       valueType: ModuleIOValueTypeEnum.string,
       label: '提取要求描述',
-      description: '给AI一些对应的背景知识或要求描述，引导AI更好的完成任务',
+      description:
+        '给AI一些对应的背景知识或要求描述，引导AI更好的完成任务。\n该输入框可使用全局变量。',
       required: true,
       placeholder:
-        '例如: \n1. 你是一个实验室预约助手，你的任务是帮助用户预约实验室。\n2. 你是谷歌搜索助手，需要从文本中提取出合适的搜索词。',
+        '例如: \n1. 当前时间为: {{cTime}}。你是一个实验室预约助手，你的任务是帮助用户预约实验室，从文本中获取对应的预约信息。\n2. 你是谷歌搜索助手，需要从文本中提取出合适的搜索词。',
       showTargetInApp: true,
       showTargetInPlugin: true
     },

packages/global/package.json

@@ -2,11 +2,12 @@
   "name": "@fastgpt/global",
   "version": "1.0.0",
   "dependencies": {
+    "axios": "^1.5.1",
     "dayjs": "^1.11.7",
-    "openai": "4.23.0",
     "encoding": "^0.1.13",
     "js-tiktoken": "^1.0.7",
-    "axios": "^1.5.1",
+    "openai": "4.23.0",
+    "pdfjs-dist": "^4.0.269",
     "timezones-list": "^3.0.2"
   },
   "devDependencies": {