update

.gitignore

@@ -1839,6 +1839,7 @@ packages/lib/services/ocr/drivers/OcrDriverTranscribe.js
 packages/lib/services/ocr/utils/createAccessiblePdf.js
 packages/lib/services/ocr/utils/filterOcrText.test.js
 packages/lib/services/ocr/utils/filterOcrText.js
+packages/lib/services/ocr/utils/glyphLessFont.js
 packages/lib/services/ocr/utils/types.js
 packages/lib/services/plugins/BasePlatformImplementation.js
 packages/lib/services/plugins/BasePluginRunner.js

.ignore.eslint

@@ -1865,6 +1865,7 @@ packages/lib/services/ocr/drivers/OcrDriverTranscribe.js
 packages/lib/services/ocr/utils/createAccessiblePdf.js
 packages/lib/services/ocr/utils/filterOcrText.test.js
 packages/lib/services/ocr/utils/filterOcrText.js
+packages/lib/services/ocr/utils/glyphLessFont.js
 packages/lib/services/ocr/utils/types.js
 packages/lib/services/plugins/BasePlatformImplementation.js
 packages/lib/services/plugins/BasePluginRunner.js

packages/lib/services/ocr/OcrService.test.ts

@@ -397,6 +397,33 @@ describe('OcrService', () => {
 		expect(pdfBytes.length).toBeGreaterThan(jpegBuffer.length * 1.5);
 	});
 
+	it.each([
+		{ words: ['ı', 'Мир', '日本語'], rtl: false },
+		{ words: ['אור', 'חדש'], rtl: true },
+		{ words: ['نور', 'باب', 'عام2024'], rtl: true },
+	])('should create an accessible PDF with text that can be extracted: $words', async ({ words, rtl }) => {
+		const jpegBuffer = await fs.readFile(`${supportDir}/photo.jpg`);
+
+		// Like Tesseract, words are in logical order, so right-to-left words are laid out from the right
+		const ocrDetails: PdfOcrDetails = {
+			version: 1,
+			pages: [{
+				lines: [{ words: words.map((t, i) => {
+					const x = rtl ? 700 - (i + 1) * 150 : 20 + i * 150;
+					return { t, bb: [x, x + 130, 20, 60] };
+				}) }],
+			}],
+		};
+
+		const pdfBytes = await createAccessiblePdf([{ buffer: jpegBuffer, width: 800, height: 400 }], JSON.stringify(ocrDetails));
+		const pdfPath = join(Setting.value('tempDir'), 'accessible.pdf');
+		await fs.writeFile(pdfPath, pdfBytes);
+
+		const textByPage = await shim.pdfExtractEmbeddedText(pdfPath);
+		expect(textByPage.length).toBe(1);
+		expect(textByPage[0].split(/\s+/)).toEqual(words);
+	});
+
 	it('should truncate long OCR text', async () => {
 		const service = newOcrService();
 		service.testing_setOcrMaxSize(10);

packages/lib/services/ocr/utils/createAccessiblePdf.ts

@@ -1,60 +1,188 @@
-import { PDFDocument, PDFFont, PDFPage, rgb, StandardFonts } from 'pdf-lib';
+import { beginText, endText, PDFDocument, PDFHexString, PDFOperator, PDFPage, PDFRef, PDFString, setCharacterSqueeze, setFontAndSize, setTextMatrix, setTextRenderingMode, showText, TextRenderingMode } from 'pdf-lib';
 import { PdfOcrDetails, RecognizeResultLine } from './types';
+import glyphLessFontBase64 from './glyphLessFont';
 
 // The PDF OCR images are created at 2x scale by pdfToImages()
 const OCR_SCALE_FACTOR = 2;
 
+// Metrics of the glyphless font, in 1/1000 of the font size
+const GLYPH_WIDTH = 500;
+const FONT_ASCENT = 800;
+const FONT_DESCENT = -200;
+
 export interface PageImageWithDimensions {
 	buffer: Buffer;
 	width: number;
 	height: number;
 }
 
+// Maps each UTF-16 code unit (used as CID) to itself. A bfrange may only vary in its last byte and a
+// block may contain at most 100 entries.
+const makeToUnicodeCMap = () => {
+	const ranges: string[] = [];
+	for (let i = 0; i < 256; i++) {
+		const high = i.toString(16).toUpperCase().padStart(2, '0');
+		ranges.push(`<${high}00> <${high}FF> <${high}00>`);
+	}
+
+	// cSpell:disable
+	const blocks: string[] = [];
+	for (let i = 0; i < ranges.length; i += 100) {
+		const chunk = ranges.slice(i, i + 100);
+		blocks.push(`${chunk.length} beginbfrange\n${chunk.join('\n')}\nendbfrange`);
+	}
+
+	return [
+		'/CIDInit /ProcSet findresource begin',
+		'12 dict begin',
+		'begincmap',
+		'/CIDSystemInfo << /Registry (Adobe) /Ordering (UCS) /Supplement 0 >> def',
+		'/CMapName /Adobe-Identity-UCS def',
+		'/CMapType 2 def',
+		'1 begincodespacerange',
+		'<0000> <FFFF>',
+		'endcodespacerange',
+		...blocks,
+		'endcmap',
+		'CMapName currentdict /CMap defineresource pop',
+		'end',
+		'end',
+	].join('\n');
+	// cSpell:enable
+};
+
+// Standard fonts only support WinAnsi, so any other character (Cyrillic, CJK, etc.) would make
+// pdf-lib throw. Instead, like Tesseract, we use a Type0 font whose CIDs are UTF-16 code units, all
+// mapped to the same empty glyph, with a ToUnicode CMap so that the text can be extracted.
+const embedGlyphLessFont = (pdfDoc: PDFDocument) => {
+	const context = pdfDoc.context;
+
+	const fontFile = Buffer.from(glyphLessFontBase64, 'base64');
+	const fontFileRef = context.register(context.flateStream(fontFile, { Length1: fontFile.length }));
+
+	// Two bytes per CID, all pointing to glyph 1
+	const cidToGidMap = new Uint8Array(0x10000 * 2);
+	for (let i = 1; i < cidToGidMap.length; i += 2) cidToGidMap[i] = 1;
+	const cidToGidMapRef = context.register(context.flateStream(cidToGidMap));
+
+	const fontDescriptorRef = context.register(context.obj({
+		Type: 'FontDescriptor',
+		FontName: 'GlyphLessFont',
+		Flags: 5,
+		FontBBox: [0, FONT_DESCENT, GLYPH_WIDTH, FONT_ASCENT],
+		ItalicAngle: 0,
+		Ascent: FONT_ASCENT,
+		Descent: FONT_DESCENT,
+		CapHeight: FONT_ASCENT,
+		StemV: 80,
+		FontFile2: fontFileRef,
+	}));
+
+	const cidFontRef = context.register(context.obj({
+		Type: 'Font',
+		Subtype: 'CIDFontType2',
+		BaseFont: 'GlyphLessFont',
+		CIDSystemInfo: {
+			Registry: PDFString.of('Adobe'),
+			Ordering: PDFString.of('Identity'),
+			Supplement: 0,
+		},
+		FontDescriptor: fontDescriptorRef,
+		CIDToGIDMap: cidToGidMapRef,
+		DW: GLYPH_WIDTH,
+	}));
+
+	const toUnicodeRef = context.register(context.flateStream(makeToUnicodeCMap()));
+
+	return context.register(context.obj({
+		Type: 'Font',
+		Subtype: 'Type0',
+		BaseFont: 'GlyphLessFont',
+		Encoding: 'Identity-H',
+		DescendantFonts: [cidFontRef],
+		ToUnicode: toUnicodeRef,
+	}));
+};
+
+const rtlScriptRegex = /[\p{Script=Hebrew}\p{Script=Arabic}\p{Script=Syriac}\p{Script=Thaana}\p{Script=Nko}]/u;
+const isRtlLetter = (char: string) => rtlScriptRegex.test(char) && !/\p{N}/u.test(char);
+const isLtrStrong = (char: string) => /[\p{L}\p{N}]/u.test(char) && !isRtlLetter(char);
+
+// OCR text is in logical order, but viewers expect PDF text in visual (left-to-right) order and
+// reorder it when extracting. So in right-to-left words, the runs are reversed, as well as the
+// characters within right-to-left runs (numbers and Latin text remain left-to-right).
+const toVisualOrder = (text: string) => {
+	const chars = Array.from(text);
+	if (!chars.some(isRtlLetter)) return text;
+
+	const runs: string[][] = [];
+	for (const char of chars) {
+		const lastRun = runs[runs.length - 1];
+		if (lastRun && isLtrStrong(lastRun[0]) === isLtrStrong(char)) {
+			lastRun.push(char);
+		} else {
+			runs.push([char]);
+		}
+	}
+
+	return runs.reverse().map(run => isLtrStrong(run[0]) ? run.join('') : run.reverse().join('')).join('');
+};
+
+const toUtf16Hex = (text: string) => {
+	let output = '';
+	for (let i = 0; i < text.length; i++) {
+		output += text.charCodeAt(i).toString(16).padStart(4, '0');
+	}
+	return output;
+};
+
 // Adds an invisible text layer to a PDF page based on OCR word positions.
 // The text is rendered in "invisible" mode (render mode 3) so it doesn't
 // appear visually but can be selected, copied, and read by screen readers.
 const addInvisibleTextLayer = (
 	page: PDFPage,
 	lines: RecognizeResultLine[],
-	font: PDFFont,
-	_pageWidth: number,
+	fontRef: PDFRef,
 	pageHeight: number,
-): void => {
+) => {
+	const fontName = page.node.newFontDictionary('GlyphLessFont', fontRef);
+	const operators: PDFOperator[] = [beginText(), setTextRenderingMode(TextRenderingMode.Invisible)];
+
 	for (const line of lines) {
 		for (const word of line.words) {
 			const text = word.t;
 			if (!text || !text.trim()) continue;
 
 			// Bounding box format from Tesseract: [x0, x1, y0, y1]
-			const [x0, , y0, y1] = word.bb;
+			const [x0, x1, y0, y1] = word.bb;
 
 			// Convert from OCR coordinates (2x scale, origin top-left)
 			// to PDF coordinates (1x scale, origin bottom-left)
-			const pdfX = x0 / OCR_SCALE_FACTOR;
-			const pdfY = pageHeight - (y1 / OCR_SCALE_FACTOR); // Flip Y axis and use bottom of bbox
-
-			// Calculate word height in PDF coordinates (width not currently used)
+			const wordWidth = (x1 - x0) / OCR_SCALE_FACTOR;
 			const wordHeight = (y1 - y0) / OCR_SCALE_FACTOR;
 
-			// Estimate font size based on word height
-			// We want the text to fit within the bounding box
-			const fontSize = Math.max(1, wordHeight * 0.85);
+			// The font's ascent-to-descent height is one em, so this makes the glyphs span the word's
+			// bounding box, with the baseline placed above its bottom by the descent.
+			const fontSize = Math.max(1, wordHeight);
+			const pdfX = x0 / OCR_SCALE_FACTOR;
+			const pdfY = pageHeight - (y1 / OCR_SCALE_FACTOR) - fontSize * FONT_DESCENT / 1000;
 
-			// Draw the text as invisible (using transparent color)
-			// PDF.js and screen readers can still detect and read this text
-			page.drawText(text, {
-				x: pdfX,
-				y: pdfY,
-				size: fontSize,
-				font: font,
-				color: rgb(0, 0, 0),
-				opacity: 0, // Make text invisible
-				// Note: pdf-lib doesn't directly support render mode 3 (invisible),
-				// but opacity: 0 achieves the same visual effect while keeping
-				// the text selectable
-			});
+			// Stretch the text horizontally so that its selection matches the word on the image
+			const hex = toUtf16Hex(toVisualOrder(text));
+			const naturalWidth = (hex.length / 4) * fontSize * GLYPH_WIDTH / 1000;
+			const horizontalScale = wordWidth > 0 ? wordWidth / naturalWidth * 100 : 100;
+
+			operators.push(
+				setFontAndSize(fontName, fontSize),
+				setCharacterSqueeze(horizontalScale),
+				setTextMatrix(1, 0, 0, 1, pdfX, pdfY),
+				showText(PDFHexString.of(hex)),
+			);
 		}
 	}
+
+	operators.push(endText());
+	page.pushOperators(...operators);
 };
 
 // Creates an accessible PDF by overlaying invisible text on top of page images.
@@ -78,8 +206,7 @@ const createAccessiblePdf = async (
 
 	const pdfDoc = await PDFDocument.create();
 
-	// Embed a standard font for the invisible text layer
-	const font = await pdfDoc.embedFont(StandardFonts.Helvetica);
+	const fontRef = embedGlyphLessFont(pdfDoc);
 
 	for (let pageIndex = 0; pageIndex < pageImages.length; pageIndex++) {
 		const pageImage = pageImages[pageIndex];
@@ -104,7 +231,7 @@ const createAccessiblePdf = async (
 		});
 
 		// Add invisible text layer on top
-		addInvisibleTextLayer(page, pageOcr.lines, font, pageWidth, pageHeight);
+		addInvisibleTextLayer(page, pageOcr.lines, fontRef, pageHeight);
 	}
 
 	return pdfDoc.save();

packages/lib/services/ocr/utils/glyphLessFont.ts

@@ -0,0 +1,7 @@
+// Minimal TrueType font generated with fontTools' FontBuilder. Same idea as Tesseract's GlyphLessFont,
+// used for invisible OCR text. Glyph 1 is a 500x1000 rectangle, with ascent 800 and descent -200.
+// It's never drawn (render mode 3), but macOS PDFKit (Preview) needs a non-zero descent and a glyph
+// outline to compute text bounds, otherwise the text can't be selected.
+const glyphLessFontBase64 = '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';
+
+export default glyphLessFontBase64;

packages/tools/cspell/dictionary4.txt

@@ -378,3 +378,7 @@ instr
 mixmark
 Hrvatski
 Defense
+glyphless
+bfrange
+flate
+Thaana