@customize-agent/knowledge 4.0.17 → 4.0.18

package/dist/extraction/content-extractor.d.ts

@@ -60,6 +60,8 @@ export declare class ContentExtractor {
     private cleanOcrText;
     /** 加载图片像素数据（依赖 sharp） */
     private loadImagePixels;
+    private readImageDimensions;
+    private isTooSmallForOcr;
     private extractPdfText;
     private normalizedTextLength;
     private toPdfTextItem;

package/dist/extraction/content-extractor.js

@@ -918,6 +918,12 @@ export class ContentExtractor {
         let imageData;
         try {
             imageData = await this.loadImagePixels(file.absolutePath);
+            if (this.isTooSmallForOcr(imageData.width, imageData.height)) {
+                metadata.contentCoverage = 'image_too_small_for_ocr';
+                metadata.imageWidth = imageData.width;
+                metadata.imageHeight = imageData.height;
+                return { text: this.metadataOnlyText(file), metadata, warnings: [`图片尺寸过小（${imageData.width}x${imageData.height}），已跳过 OCR 并仅索引元数据`] };
+            }
         }
         catch (e) {
             metadata.contentCoverage = 'image_decode_failed';
@@ -1094,10 +1100,16 @@ export class ContentExtractor {
         for (let i = 0; i < pageImages.length; i++) {
             const imgPath = pageImages[i];
             try {
+                const dimensions = await this.readImageDimensions(imgPath);
+                if (dimensions && this.isTooSmallForOcr(dimensions.width, dimensions.height)) {
+                    failedPages.push({ page: i + 1, reason: `image_too_small_for_ocr_${dimensions.width}x${dimensions.height}` });
+                    warnings.push(`PDF 第 ${i + 1} 页渲染图片尺寸过小（${dimensions.width}x${dimensions.height}），已跳过 OCR`);
+                    continue;
+                }
                 const provider = await getOcrProvider();
                 const ocrResult = await provider.recognize({
                     data: new Uint8Array(0),
-                    width: 0, height: 0, channels: 0,
+                    width: dimensions?.width ?? 0, height: dimensions?.height ?? 0, channels: 0,
                     filePath: imgPath,
                 });
                 const ocrText = this.cleanOcrText(ocrResult.text);
@@ -1219,6 +1231,22 @@ export class ContentExtractor {
         const { data, info } = await sharpFn(filePath).raw().toBuffer({ resolveWithObject: true });
         return { data: new Uint8Array(data), width: info.width, height: info.height };
     }
+    async readImageDimensions(filePath) {
+        try {
+            const sharpMod = await resolveAndImport('sharp');
+            const sharpFn = sharpMod.default ?? sharpMod;
+            const metadata = await sharpFn(filePath).metadata();
+            const width = Number(metadata.width ?? 0);
+            const height = Number(metadata.height ?? 0);
+            return width > 0 && height > 0 ? { width, height } : undefined;
+        }
+        catch {
+            return undefined;
+        }
+    }
+    isTooSmallForOcr(width, height) {
+        return width < 8 || height < 8 || width * height < 128;
+    }
     async extractPdfText(buffer) {
         let pdfjsText = '';
         // 第一层：pdfjs-dist 文本提取（处理压缩内容流、CJK 字体、现代 PDF）

package/dist/extraction/ocr-providers.d.ts

@@ -48,6 +48,8 @@ export declare class TesseractJsProvider implements OcrProvider {
         filePath?: string;
     }): Promise<OcrResult>;
     getWarnings(): string[];
+    private readImageDimensions;
+    private isTooSmallForOcr;
     private getWorker;
     private createReusableWorker;
     dispose(): Promise<void>;

package/dist/extraction/ocr-providers.js

@@ -44,9 +44,14 @@ export class TesseractJsProvider {
     async recognize(input) {
         let pngPath;
         let tmpDir = null;
+        let width = input.width;
+        let height = input.height;
         // 如果传了 filePath，直接使用；否则 raw pixels → PNG
         if (input.filePath && fs.existsSync(input.filePath)) {
             pngPath = input.filePath;
+            const dimensions = await this.readImageDimensions(pngPath);
+            width = dimensions?.width ?? width;
+            height = dimensions?.height ?? height;
         }
         else {
             const sharpMod = await resolveAndImport('sharp');
@@ -61,6 +66,9 @@ export class TesseractJsProvider {
                 .withMetadata({ density: 288 })
                 .png().toFile(pngPath);
         }
+        if (this.isTooSmallForOcr(width, height)) {
+            return { text: '', confidence: 0, regions: [], warnings: [`image too small for OCR: ${width}x${height}`] };
+        }
         try {
             const worker = await this.getWorker();
             const result = await worker.recognize(pngPath);
@@ -88,6 +96,22 @@ export class TesseractJsProvider {
     getWarnings() {
         return [...new Set(this.warnings)].slice(-20);
     }
+    async readImageDimensions(filePath) {
+        try {
+            const sharpMod = await resolveAndImport('sharp');
+            const sharpFn = sharpMod.default ?? sharpMod;
+            const metadata = await sharpFn(filePath).metadata();
+            const width = Number(metadata.width ?? 0);
+            const height = Number(metadata.height ?? 0);
+            return width > 0 && height > 0 ? { width, height } : undefined;
+        }
+        catch {
+            return undefined;
+        }
+    }
+    isTooSmallForOcr(width, height) {
+        return width < 8 || height < 8 || width * height < 128;
+    }
     async getWorker() {
         if (this.worker)
             return this.worker;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@customize-agent/knowledge",
-  "version": "4.0.17",
+  "version": "4.0.18",
   "description": "Local knowledge base infrastructure for customize-agent",
   "type": "module",
   "main": "./dist/index.js",