@stabrise/scaledp 0.1.0

package/dist/pdf-BQl0dneD.js

@@ -0,0 +1,417 @@
+import { c as ImageError, h as getConfig, i as Stage } from "./pipeline-DACqGkpN.js";
+import { c as encodeImage, i as createCanvas, r as context2d } from "./image-CAH2rLv9.js";
+import { i as resolveParams, t as BASE_STAGE_DEFAULTS } from "./params-DapwK9Ns.js";
+import { n as createDocument, t as createImage } from "./image-DoZDJkcR.js";
+import { r as toBytes } from "./data-to-image-DoZ4jQ3R.js";
+//#region src/pdf/extract-text.ts
+/** Glyphs sit above the baseline by roughly three quarters of the line height. */
+const ASCENT_RATIO = .75;
+/** Confidence assigned to text read from a PDF's own text layer. */
+const TEXT_LAYER_SCORE = .99;
+function isTextItem(item) {
+	return typeof item === "object" && item !== null && "str" in item && "transform" in item && Array.isArray(item.transform);
+}
+/** Convert one pdf.js text item into a viewport-space box. */
+function textItemToBox(item, viewport) {
+	if (item.str.length === 0) return null;
+	const [a = 0, b = 0, c = 0, d = 0, e = 0, f = 0] = item.transform;
+	const abMag = Math.hypot(a, b) || 1;
+	const cdMag = Math.hypot(c, d) || 1;
+	const dirX = a / abMag;
+	const dirY = b / abMag;
+	const upX = c / cdMag;
+	const upY = d / cdMag;
+	const ascent = item.height * ASCENT_RATIO;
+	const startX = e + upX * ascent;
+	const startY = f + upY * ascent;
+	const corners = [
+		viewport.convertToViewportPoint(startX, startY),
+		viewport.convertToViewportPoint(startX + dirX * item.width, startY + dirY * item.width),
+		viewport.convertToViewportPoint(startX - upX * item.height, startY - upY * item.height),
+		viewport.convertToViewportPoint(startX + dirX * item.width - upX * item.height, startY + dirY * item.width - upY * item.height)
+	];
+	const xs = corners.map((p) => p[0]);
+	const ys = corners.map((p) => p[1]);
+	const x = Math.min(...xs);
+	const y = Math.min(...ys);
+	const [startScreenX = 0, startScreenY = 0] = corners[0] ?? [];
+	const [endScreenX = 0, endScreenY = 0] = corners[1] ?? [];
+	const readMag = Math.hypot(endScreenX - startScreenX, endScreenY - startScreenY) || 1;
+	return {
+		text: item.str,
+		score: TEXT_LAYER_SCORE,
+		x: Math.floor(x),
+		y: Math.floor(y),
+		width: Math.max(1, Math.ceil(Math.max(...xs) - x)),
+		height: Math.max(1, Math.ceil(Math.max(...ys) - y)),
+		angle: 0,
+		readDirX: (endScreenX - startScreenX) / readMag,
+		readDirY: (endScreenY - startScreenY) / readMag,
+		fontName: item.fontName ?? ""
+	};
+}
+/** Extract every text item on a page as a viewport-space box. */
+async function extractTextBoxes(page, viewport) {
+	const content = await page.getTextContent();
+	const boxes = [];
+	for (const item of content.items) {
+		if (!isTextItem(item)) continue;
+		const box = textItemToBox(item, viewport);
+		if (box) boxes.push(box);
+	}
+	return boxes;
+}
+//#endregion
+//#region src/pdf/pdfjs.ts
+/**
+* Lazy pdf.js loader.
+*
+* pdfjs-dist is an optional peer dependency, so it is imported only when a PDF
+* stage actually runs. Every asset path comes from `configure()` -- unlike the
+* pdftools prototype, which hardcoded `/pdf.worker.min.mjs`, a path only its
+* own Next app could serve.
+*/
+let modulePromise = null;
+async function loadPdfjs() {
+	if (modulePromise) return modulePromise;
+	modulePromise = (async () => {
+		let pdfjs;
+		try {
+			pdfjs = await import("pdfjs-dist");
+		} catch (cause) {
+			throw new Error("pdfjs-dist is required for PDF support. Install it: npm i pdfjs-dist", { cause });
+		}
+		const { workerSrc } = getConfig().pdf;
+		if (workerSrc) pdfjs.GlobalWorkerOptions.workerSrc = workerSrc;
+		return pdfjs;
+	})();
+	return modulePromise;
+}
+/** Reset the cached module. Tests only. */
+function resetPdfjs() {
+	modulePromise = null;
+}
+/**
+* Turn pdf.js's worker-setup failure into something actionable.
+*
+* When `workerSrc` is unset or 404s, pdf.js reports "Setting up fake worker
+* failed" with a bare module URL, which says nothing about what to do. The
+* worker is not bundled with this library on purpose -- it has to be served by
+* the consuming application -- so the fix is always the same two steps.
+*/
+function describePdfError(error) {
+	const message = error instanceof Error ? error.message : String(error);
+	if (!/fake worker|worker/i.test(message)) return error instanceof Error ? error : new Error(message);
+	const { workerSrc } = getConfig().pdf;
+	const cause = workerSrc ? `pdf.js could not load its worker from "${workerSrc}".` : "pdf.js has no worker configured.";
+	return new Error(`${cause}\nCopy it out of the package and point the config at it:
+  cp node_modules/pdfjs-dist/build/pdf.worker.min.mjs public/
+  configure({ pdf: { workerSrc: '/pdf.worker.min.mjs' } })
+Original error: ${message}`, { cause: error });
+}
+/** Document-level options assembled from the global config. */
+function documentOptions(data) {
+	const { cMapUrl, standardFontDataUrl, wasmUrl } = getConfig().pdf;
+	const owned = new Uint8Array(data.byteLength);
+	owned.set(data);
+	const options = { data: owned };
+	if (cMapUrl) {
+		options.cMapUrl = cMapUrl;
+		options.cMapPacked = true;
+	}
+	if (standardFontDataUrl) options.standardFontDataUrl = standardFontDataUrl;
+	if (wasmUrl) options.wasmUrl = wasmUrl;
+	return options;
+}
+//#endregion
+//#region src/pdf/pdf-to-image.ts
+/**
+* Port of `scaledp/pdf/PdfDataToImage.py`: a PDF into one `Image` row per page.
+*
+* Python renders with PyMuPDF at a DPI; pdf.js works in scale factors, so the
+* DPI converts through the PDF unit of 72 points per inch.
+*/
+/** PDF user space is defined in points; 72 of them make an inch. */
+const POINTS_PER_INCH = 72;
+const PDF_TO_IMAGE_DEFAULTS = Object.freeze({
+	...BASE_STAGE_DEFAULTS,
+	inputCol: "content",
+	outputCol: "image",
+	resolution: 300,
+	pageLimit: 0,
+	imageType: "png"
+});
+const MIME = {
+	png: "image/png",
+	webp: "image/webp",
+	jpeg: "image/jpeg"
+};
+var PdfToImage = class extends Stage {
+	name = "PdfToImage";
+	constructor(options = {}) {
+		super(resolveParams(PDF_TO_IMAGE_DEFAULTS, options, {
+			resolution: (value) => {
+				if (!Number.isFinite(value) || value <= 0) throw new RangeError(`resolution must be positive, received ${value}`);
+			},
+			pageLimit: (value) => {
+				if (!Number.isInteger(value) || value < 0) throw new RangeError(`pageLimit must be a non-negative integer, received ${value}`);
+			}
+		}));
+	}
+	/** One input PDF becomes N rows, each carrying its page index. */
+	async expand(input, row, ctx) {
+		const { outputCol, pageCol, pathCol, resolution, pageLimit, imageType } = this.params;
+		const path = String(row[pathCol] ?? "memory");
+		const task = (await loadPdfjs()).getDocument(documentOptions(toBytes(input)));
+		try {
+			const document = await task.promise;
+			const pageCount = pageLimit > 0 ? Math.min(pageLimit, document.numPages) : document.numPages;
+			const rows = [];
+			for (let index = 0; index < pageCount; index++) {
+				ctx.signal?.throwIfAborted();
+				const image = await renderPage(document, index + 1, {
+					resolution,
+					imageType,
+					path
+				});
+				rows.push({
+					...row,
+					[pageCol]: index,
+					[outputCol]: image
+				});
+			}
+			return rows;
+		} catch (error) {
+			throw describePdfError(error);
+		} finally {
+			await task.destroy();
+		}
+	}
+	async apply() {
+		throw new ImageError("unreachable: expand handles every row", this.name);
+	}
+	onError(message, row) {
+		return createImage({
+			path: String(row[this.params.pathCol] ?? "memory"),
+			exception: message
+		});
+	}
+};
+/** Rasterise a single 1-based page to encoded image bytes. */
+async function renderPage(document, pageNumber, opts) {
+	const page = await document.getPage(pageNumber);
+	try {
+		const viewport = page.getViewport({ scale: opts.resolution / 72 });
+		const canvas = createCanvas(viewport.width, viewport.height);
+		await page.render({
+			canvas,
+			viewport
+		}).promise;
+		return createImage({
+			path: opts.path,
+			resolution: opts.resolution,
+			data: await encodeImage(canvas, MIME[opts.imageType]),
+			imageType: opts.imageType,
+			width: canvas.width,
+			height: canvas.height
+		});
+	} finally {
+		page.cleanup();
+	}
+}
+//#endregion
+//#region src/pdf/split-words.ts
+/**
+* Split a pdf.js text run into word-level boxes.
+*
+* pdf.js emits runs at line granularity ("Client: Raja Raman"), but detection
+* boxes and NER offsets both want words. Each word is measured with Canvas 2D
+* `measureText` using a font reconstructed from the pdf.js font name, then the
+* measured widths are scaled so they sum to the run's actual width -- the
+* substitute font is never metrically identical to the embedded one, so the
+* measurements are only useful as *proportions*.
+*
+* Walking along `readDirX`/`readDirY` rather than assuming left-to-right is what
+* makes rotated, bottom-to-top and right-to-left runs come out correctly.
+*/
+/** Widen each word slightly so glyph overhang is not clipped. */
+const WORD_PADDING_RATIO = .1;
+let measureCtx = null;
+function measurementContext() {
+	if (measureCtx) return measureCtx;
+	try {
+		measureCtx = context2d(createCanvas(1, 1));
+		return measureCtx;
+	} catch {
+		return null;
+	}
+}
+/**
+* Rebuild a CSS font string from a pdf.js font name.
+*
+* Names look like `AAAAAA+Helvetica-BoldOblique`: a six-letter subset prefix,
+* then the real family and style suffixes.
+*/
+function cssFontFromPdfName(fontName, size) {
+	const lower = fontName.replace(/^[A-Z]{6}\+/, "").toLowerCase();
+	const weight = /bold|black|heavy|semibold/.test(lower) ? "bold" : "normal";
+	const style = /italic|oblique/.test(lower) ? "italic" : "normal";
+	const family = /serif|times|georgia|garamond|roman/.test(lower) ? "serif" : /mono|courier|consol/.test(lower) ? "monospace" : "sans-serif";
+	return `${style} ${weight} ${Math.max(1, Math.round(size))}px ${family}`;
+}
+/**
+* Relative advance width per character, used when no canvas is available.
+* Buckets rather than real metrics -- enough to keep proportions sane.
+*/
+function relativeCharWidth(char) {
+	if ("iljI|.,:;'`!".includes(char)) return .6;
+	if ("ftr()[]{}-".includes(char)) return .8;
+	if ("MWmw@%".includes(char)) return 1.6;
+	if (char === " ") return .6;
+	if (char >= "A" && char <= "Z") return 1.3;
+	return 1;
+}
+function measureWord(word, font) {
+	const ctx = measurementContext();
+	if (ctx) {
+		ctx.font = font;
+		return ctx.measureText(word).width;
+	}
+	let total = 0;
+	for (const char of word) total += relativeCharWidth(char);
+	return total;
+}
+/**
+* Split one run into word boxes.
+*
+* Returns the run itself when it holds a single word, so the common case costs
+* nothing.
+*/
+function splitRunIntoWords(run) {
+	const trimmed = run.text.trim();
+	if (trimmed.length === 0) return [];
+	const tokens = trimmed.split(/(\s+)/).filter((t) => t.length > 0);
+	if (tokens.filter((t) => !/^\s+$/.test(t)).length <= 1) return [{
+		...run,
+		text: trimmed
+	}];
+	const font = cssFontFromPdfName(run.fontName, run.height);
+	const measured = tokens.map((token) => measureWord(token, font));
+	const totalMeasured = measured.reduce((sum, w) => sum + w, 0) || 1;
+	const scale = (Math.hypot(run.width * run.readDirX, run.height * run.readDirY) || run.width) / totalMeasured;
+	let cursorX = run.readDirX >= 0 ? run.x : run.x + run.width;
+	let cursorY = run.readDirY >= 0 ? run.y : run.y + run.height;
+	const out = [];
+	for (const [i, token] of tokens.entries()) {
+		const advance = measured[i] * scale;
+		if (!/^\s+$/.test(token)) {
+			const pad = run.height * WORD_PADDING_RATIO;
+			const spanX = Math.abs(run.readDirX) > Math.abs(run.readDirY) ? advance : run.width;
+			const spanY = Math.abs(run.readDirY) > Math.abs(run.readDirX) ? advance : run.height;
+			const left = run.readDirX >= 0 ? cursorX : cursorX - spanX;
+			const top = run.readDirY >= 0 ? cursorY : cursorY - spanY;
+			out.push({
+				text: token,
+				score: run.score,
+				x: Math.floor(left - pad),
+				y: Math.floor(top - pad),
+				width: Math.max(1, Math.ceil(spanX + pad * 2)),
+				height: Math.max(1, Math.ceil(spanY + pad * 2)),
+				angle: run.angle
+			});
+		}
+		cursorX += run.readDirX * advance;
+		cursorY += run.readDirY * advance;
+	}
+	return out;
+}
+/** Split every run on a page into word boxes. */
+function splitRunsIntoWords(runs) {
+	return runs.flatMap(splitRunIntoWords);
+}
+/** Reset the cached measurement canvas. Tests only. */
+function resetMeasurementContext() {
+	measureCtx = null;
+}
+//#endregion
+//#region src/pdf/pdf-to-document.ts
+/**
+* Port of `scaledp/pdf/PdfDataToText.py`: a PDF's embedded text layer into one
+* `Document` row per page, with word-level boxes.
+*
+* Coordinates are emitted in the same pixel space `PdfToImage` renders at, so
+* boxes from this stage and boxes from OCR are directly comparable. Python
+* leaves PdfDataToText in PDF points and scales only in PdfDataToDocument; a
+* single consistent space is more useful and avoids a class of silent mismatch.
+*
+* The output feeds the `bypassCol` optimisation: a page that already has a text
+* layer does not need OCR.
+*/
+const PDF_TO_DOCUMENT_DEFAULTS = Object.freeze({
+	...BASE_STAGE_DEFAULTS,
+	inputCol: "content",
+	outputCol: "document",
+	keepInputData: true,
+	resolution: 300,
+	pageLimit: 0,
+	splitWords: true
+});
+var PdfToDocument = class extends Stage {
+	name = "PdfToDocument";
+	constructor(options = {}) {
+		super(resolveParams(PDF_TO_DOCUMENT_DEFAULTS, options));
+	}
+	async expand(input, row, ctx) {
+		const { outputCol, pageCol, pathCol, resolution, pageLimit, splitWords } = this.params;
+		const path = String(row[pathCol] ?? "memory");
+		const task = (await loadPdfjs()).getDocument(documentOptions(toBytes(input)));
+		try {
+			const pdf = await task.promise;
+			const pageCount = pageLimit > 0 ? Math.min(pageLimit, pdf.numPages) : pdf.numPages;
+			const rows = [];
+			for (let index = 0; index < pageCount; index++) {
+				ctx.signal?.throwIfAborted();
+				const page = await pdf.getPage(index + 1);
+				try {
+					const runs = await extractTextBoxes(page, page.getViewport({ scale: resolution / 72 }));
+					const bboxes = splitWords ? splitRunsIntoWords(runs) : runs;
+					rows.push({
+						...row,
+						[pageCol]: index,
+						[outputCol]: createDocument({
+							path,
+							type: "pdf",
+							text: runs.map((r) => r.text).join("\n"),
+							bboxes
+						})
+					});
+				} finally {
+					page.cleanup();
+				}
+			}
+			return rows;
+		} catch (error) {
+			throw describePdfError(error);
+		} finally {
+			await task.destroy();
+		}
+	}
+	async apply() {
+		throw new ImageError("unreachable: expand handles every row", this.name);
+	}
+	onError(message, row) {
+		return createDocument({
+			path: String(row[this.params.pathCol] ?? "memory"),
+			type: "pdf",
+			exception: message
+		});
+	}
+};
+/** True when a page's text layer is substantive enough to skip OCR. */
+function hasUsableTextLayer(document, minimumBoxes = 1) {
+	return document.exception === "" && document.bboxes.length >= minimumBoxes;
+}
+//#endregion
+export { extractTextBoxes as _, relativeCharWidth as a, splitRunsIntoWords as c, PdfToImage as d, renderPage as f, TEXT_LAYER_SCORE as g, resetPdfjs as h, cssFontFromPdfName as i, PDF_TO_IMAGE_DEFAULTS as l, loadPdfjs as m, PdfToDocument as n, resetMeasurementContext as o, documentOptions as p, hasUsableTextLayer as r, splitRunIntoWords as s, PDF_TO_DOCUMENT_DEFAULTS as t, POINTS_PER_INCH as u, isTextItem as v, textItemToBox as y };
+
+//# sourceMappingURL=pdf-BQl0dneD.js.map

package/dist/pdf-BQl0dneD.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"pdf-BQl0dneD.js","names":[],"sources":["../src/pdf/extract-text.ts","../src/pdf/pdfjs.ts","../src/pdf/pdf-to-image.ts","../src/pdf/split-words.ts","../src/pdf/pdf-to-document.ts"],"sourcesContent":["/**\n * Word-level text extraction from a PDF's embedded text layer.\n *\n * Ported from the pdftools prototype's `extractTextFromPage`, whose rotation\n * handling is the hard-won part and is reproduced here with its reasoning.\n *\n * A pdf.js text item carries a full transform matrix [a, b, c, d, e, f]. Naively\n * reading `a` as a width scale breaks on rotated text, so the reading direction\n * and the \"up\" direction are recovered as *unit* vectors -- otherwise the font\n * size gets counted twice. The four corners are then pushed through\n * `convertToViewportPoint` and reduced to an axis-aligned box.\n *\n * No semantic `angle` is emitted, deliberately. A rotated glyph matrix is\n * indistinguishable from page-level /Rotate compensation or an embedded\n * FontMatrix, so composing the corners geometrically and taking their bounding\n * box is correct regardless of which caused it. `readDirX`/`readDirY` carry the\n * direction that word-splitting needs.\n */\n\nimport type { Box } from '../schemas/box.js'\n\n/** A box plus the reading direction, which splitting needs but ScaleDP's Box lacks. */\nexport interface TextBox extends Box {\n    /** Unit vector, in viewport space, along which reading progresses. */\n    readDirX: number\n    readDirY: number\n    /** pdf.js font identifier, e.g. 'g_d0_f1'. */\n    fontName: string\n}\n\n/** Glyphs sit above the baseline by roughly three quarters of the line height. */\nconst ASCENT_RATIO = 0.75\n\n/** Confidence assigned to text read from a PDF's own text layer. */\nexport const TEXT_LAYER_SCORE = 0.99\n\ninterface TextItemLike {\n    str: string\n    transform: number[]\n    width: number\n    height: number\n    fontName?: string\n}\n\ninterface ViewportLike {\n    convertToViewportPoint(x: number, y: number): number[]\n}\n\nexport function isTextItem(item: unknown): item is TextItemLike {\n    return (\n        typeof item === 'object' &&\n        item !== null &&\n        'str' in item &&\n        'transform' in item &&\n        Array.isArray((item as TextItemLike).transform)\n    )\n}\n\n/** Convert one pdf.js text item into a viewport-space box. */\nexport function textItemToBox(item: TextItemLike, viewport: ViewportLike): TextBox | null {\n    if (item.str.length === 0) return null\n\n    const [a = 0, b = 0, c = 0, d = 0, e = 0, f = 0] = item.transform\n\n    // Unit direction vectors. Dividing out the magnitudes is what stops the\n    // font size being applied twice, since item.width/height already include it.\n    const abMag = Math.hypot(a, b) || 1\n    const cdMag = Math.hypot(c, d) || 1\n    const dirX = a / abMag\n    const dirY = b / abMag\n    const upX = c / cdMag\n    const upY = d / cdMag\n\n    // The transform's origin is the baseline; shift up to the glyph tops.\n    const ascent = item.height * ASCENT_RATIO\n    const startX = e + upX * ascent\n    const startY = f + upY * ascent\n\n    const corners = [\n        viewport.convertToViewportPoint(startX, startY),\n        viewport.convertToViewportPoint(startX + dirX * item.width, startY + dirY * item.width),\n        viewport.convertToViewportPoint(startX - upX * item.height, startY - upY * item.height),\n        viewport.convertToViewportPoint(\n            startX + dirX * item.width - upX * item.height,\n            startY + dirY * item.width - upY * item.height\n        ),\n    ]\n\n    const xs = corners.map((p) => p[0] as number)\n    const ys = corners.map((p) => p[1] as number)\n    const x = Math.min(...xs)\n    const y = Math.min(...ys)\n\n    // Reading direction in viewport space, taken from the start and end points\n    // rather than from the matrix, so the viewport's own flip is accounted for.\n    const [startScreenX = 0, startScreenY = 0] = corners[0] ?? []\n    const [endScreenX = 0, endScreenY = 0] = corners[1] ?? []\n    const readMag = Math.hypot(endScreenX - startScreenX, endScreenY - startScreenY) || 1\n\n    return {\n        text: item.str,\n        score: TEXT_LAYER_SCORE,\n        x: Math.floor(x),\n        y: Math.floor(y),\n        width: Math.max(1, Math.ceil(Math.max(...xs) - x)),\n        height: Math.max(1, Math.ceil(Math.max(...ys) - y)),\n        angle: 0,\n        readDirX: (endScreenX - startScreenX) / readMag,\n        readDirY: (endScreenY - startScreenY) / readMag,\n        fontName: item.fontName ?? '',\n    }\n}\n\n/** Extract every text item on a page as a viewport-space box. */\nexport async function extractTextBoxes(\n    page: { getTextContent(): Promise<{ items: unknown[] }> },\n    viewport: ViewportLike\n): Promise<TextBox[]> {\n    const content = await page.getTextContent()\n    const boxes: TextBox[] = []\n    for (const item of content.items) {\n        if (!isTextItem(item)) continue\n        const box = textItemToBox(item, viewport)\n        if (box) boxes.push(box)\n    }\n    return boxes\n}\n","/**\n * Lazy pdf.js loader.\n *\n * pdfjs-dist is an optional peer dependency, so it is imported only when a PDF\n * stage actually runs. Every asset path comes from `configure()` -- unlike the\n * pdftools prototype, which hardcoded `/pdf.worker.min.mjs`, a path only its\n * own Next app could serve.\n */\n\nimport { getConfig } from '../core/config.js'\n\ntype PdfjsModule = typeof import('pdfjs-dist')\n\nlet modulePromise: Promise<PdfjsModule> | null = null\n\nexport async function loadPdfjs(): Promise<PdfjsModule> {\n    if (modulePromise) return modulePromise\n\n    modulePromise = (async () => {\n        let pdfjs: PdfjsModule\n        try {\n            pdfjs = await import('pdfjs-dist')\n        } catch (cause) {\n            throw new Error('pdfjs-dist is required for PDF support. Install it: npm i pdfjs-dist', { cause })\n        }\n\n        const { workerSrc } = getConfig().pdf\n        if (workerSrc) pdfjs.GlobalWorkerOptions.workerSrc = workerSrc\n        return pdfjs\n    })()\n\n    return modulePromise\n}\n\n/** Reset the cached module. Tests only. */\nexport function resetPdfjs(): void {\n    modulePromise = null\n}\n\n/**\n * Turn pdf.js's worker-setup failure into something actionable.\n *\n * When `workerSrc` is unset or 404s, pdf.js reports \"Setting up fake worker\n * failed\" with a bare module URL, which says nothing about what to do. The\n * worker is not bundled with this library on purpose -- it has to be served by\n * the consuming application -- so the fix is always the same two steps.\n */\nexport function describePdfError(error: unknown): Error {\n    const message = error instanceof Error ? error.message : String(error)\n    if (!/fake worker|worker/i.test(message)) {\n        return error instanceof Error ? error : new Error(message)\n    }\n\n    const { workerSrc } = getConfig().pdf\n    const cause = workerSrc\n        ? `pdf.js could not load its worker from \"${workerSrc}\".`\n        : 'pdf.js has no worker configured.'\n\n    return new Error(\n        `${cause}\\n` +\n            'Copy it out of the package and point the config at it:\\n' +\n            '  cp node_modules/pdfjs-dist/build/pdf.worker.min.mjs public/\\n' +\n            \"  configure({ pdf: { workerSrc: '/pdf.worker.min.mjs' } })\\n\" +\n            `Original error: ${message}`,\n        { cause: error }\n    )\n}\n\n/** Document-level options assembled from the global config. */\nexport function documentOptions(data: Uint8Array): Record<string, unknown> {\n    const { cMapUrl, standardFontDataUrl, wasmUrl } = getConfig().pdf\n    // pdf.js takes ownership of the buffer it is given and detaches it, so hand\n    // over a copy: callers routinely reuse the row's `content` afterwards.\n    const owned = new Uint8Array(data.byteLength)\n    owned.set(data)\n\n    const options: Record<string, unknown> = { data: owned }\n    if (cMapUrl) {\n        options.cMapUrl = cMapUrl\n        options.cMapPacked = true\n    }\n    if (standardFontDataUrl) options.standardFontDataUrl = standardFontDataUrl\n    if (wasmUrl) options.wasmUrl = wasmUrl\n    return options\n}\n","/**\n * Port of `scaledp/pdf/PdfDataToImage.py`: a PDF into one `Image` row per page.\n *\n * Python renders with PyMuPDF at a DPI; pdf.js works in scale factors, so the\n * DPI converts through the PDF unit of 72 points per inch.\n */\n\nimport { ImageError } from '../core/errors.js'\nimport { createCanvas, encodeImage } from '../core/image.js'\nimport { BASE_STAGE_DEFAULTS, type BaseStageParams, resolveParams } from '../core/params.js'\nimport { type Row, Stage, type StageContext } from '../core/pipeline.js'\nimport { createImage, type ImageFormat, type ScaleDpImage } from '../schemas/image.js'\nimport { toBytes } from '../stages/data-to-image.js'\nimport { describePdfError, documentOptions, loadPdfjs } from './pdfjs.js'\n\n/** PDF user space is defined in points; 72 of them make an inch. */\nexport const POINTS_PER_INCH = 72\n\nexport interface PdfToImageParams extends BaseStageParams {\n    /** Render DPI. 300 matches ScaleDP's default and suits OCR. */\n    resolution: number\n    /** Maximum pages to render; 0 renders all of them. */\n    pageLimit: number\n    imageType: ImageFormat\n}\n\nexport const PDF_TO_IMAGE_DEFAULTS: PdfToImageParams = Object.freeze({\n    ...BASE_STAGE_DEFAULTS,\n    inputCol: 'content',\n    outputCol: 'image',\n    resolution: 300,\n    pageLimit: 0,\n    imageType: 'png' as ImageFormat,\n})\n\nconst MIME: Record<ImageFormat, 'image/png' | 'image/webp' | 'image/jpeg'> = {\n    png: 'image/png',\n    webp: 'image/webp',\n    jpeg: 'image/jpeg',\n}\n\nexport class PdfToImage extends Stage<PdfToImageParams> {\n    readonly name = 'PdfToImage'\n\n    constructor(options: Partial<PdfToImageParams> = {}) {\n        super(\n            resolveParams(PDF_TO_IMAGE_DEFAULTS, options, {\n                resolution: (value) => {\n                    if (!Number.isFinite(value) || value <= 0) {\n                        throw new RangeError(`resolution must be positive, received ${value}`)\n                    }\n                },\n                pageLimit: (value) => {\n                    if (!Number.isInteger(value) || value < 0) {\n                        throw new RangeError(`pageLimit must be a non-negative integer, received ${value}`)\n                    }\n                },\n            })\n        )\n    }\n\n    /** One input PDF becomes N rows, each carrying its page index. */\n    protected override async expand(input: unknown, row: Row, ctx: StageContext): Promise<Row[]> {\n        const { outputCol, pageCol, pathCol, resolution, pageLimit, imageType } = this.params\n        const path = String(row[pathCol] ?? 'memory')\n\n        const pdfjs = await loadPdfjs()\n        const task = pdfjs.getDocument(documentOptions(toBytes(input)))\n\n        try {\n            // pdf.js defers worker setup, so a missing worker surfaces on first\n            // page access rather than from task.promise. Wrap the whole\n            // operation so the failure is described wherever it lands.\n            const document = await task.promise\n            const pageCount = pageLimit > 0 ? Math.min(pageLimit, document.numPages) : document.numPages\n            const rows: Row[] = []\n\n            for (let index = 0; index < pageCount; index++) {\n                ctx.signal?.throwIfAborted()\n                const image = await renderPage(document, index + 1, {\n                    resolution,\n                    imageType,\n                    path,\n                })\n                rows.push({ ...row, [pageCol]: index, [outputCol]: image })\n            }\n            return rows\n        } catch (error) {\n            throw describePdfError(error)\n        } finally {\n            // destroy() lives on the loading task, not the document proxy, and\n            // is what releases the pdf.js worker's copy of the file.\n            await task.destroy()\n        }\n    }\n\n    protected async apply(): Promise<never> {\n        throw new ImageError('unreachable: expand handles every row', this.name)\n    }\n\n    protected onError(message: string, row: Row): ScaleDpImage {\n        return createImage({ path: String(row[this.params.pathCol] ?? 'memory'), exception: message })\n    }\n}\n\n/** Rasterise a single 1-based page to encoded image bytes. */\nexport async function renderPage(\n    document: Awaited<ReturnType<typeof import('pdfjs-dist').getDocument>['promise']>,\n    pageNumber: number,\n    opts: { resolution: number; imageType: ImageFormat; path: string }\n): Promise<ScaleDpImage> {\n    const page = await document.getPage(pageNumber)\n    try {\n        const viewport = page.getViewport({ scale: opts.resolution / POINTS_PER_INCH })\n        const canvas = createCanvas(viewport.width, viewport.height)\n\n        // pdf.js >= 5 takes `canvas`, not `canvasContext`. Passing the context\n        // still works but is the documented legacy path.\n        await page.render({\n            canvas: canvas as unknown as HTMLCanvasElement,\n            viewport,\n        }).promise\n\n        return createImage({\n            path: opts.path,\n            resolution: opts.resolution,\n            data: await encodeImage(canvas, MIME[opts.imageType]),\n            imageType: opts.imageType,\n            width: canvas.width,\n            height: canvas.height,\n        })\n    } finally {\n        page.cleanup()\n    }\n}\n","/**\n * Split a pdf.js text run into word-level boxes.\n *\n * pdf.js emits runs at line granularity (\"Client: Raja Raman\"), but detection\n * boxes and NER offsets both want words. Each word is measured with Canvas 2D\n * `measureText` using a font reconstructed from the pdf.js font name, then the\n * measured widths are scaled so they sum to the run's actual width -- the\n * substitute font is never metrically identical to the embedded one, so the\n * measurements are only useful as *proportions*.\n *\n * Walking along `readDirX`/`readDirY` rather than assuming left-to-right is what\n * makes rotated, bottom-to-top and right-to-left runs come out correctly.\n */\n\nimport { context2d, createCanvas } from '../core/image.js'\nimport type { Box } from '../schemas/box.js'\nimport type { TextBox } from './extract-text.js'\n\n/** Widen each word slightly so glyph overhang is not clipped. */\nconst WORD_PADDING_RATIO = 0.1\n\nlet measureCtx: OffscreenCanvasRenderingContext2D | null = null\n\nfunction measurementContext(): OffscreenCanvasRenderingContext2D | null {\n    if (measureCtx) return measureCtx\n    try {\n        measureCtx = context2d(createCanvas(1, 1))\n        return measureCtx\n    } catch {\n        // No canvas (e.g. a non-browser test run): fall back to glyph heuristics.\n        return null\n    }\n}\n\n/**\n * Rebuild a CSS font string from a pdf.js font name.\n *\n * Names look like `AAAAAA+Helvetica-BoldOblique`: a six-letter subset prefix,\n * then the real family and style suffixes.\n */\nexport function cssFontFromPdfName(fontName: string, size: number): string {\n    const name = fontName.replace(/^[A-Z]{6}\\+/, '')\n    const lower = name.toLowerCase()\n    const weight = /bold|black|heavy|semibold/.test(lower) ? 'bold' : 'normal'\n    const style = /italic|oblique/.test(lower) ? 'italic' : 'normal'\n    const family = /serif|times|georgia|garamond|roman/.test(lower)\n        ? 'serif'\n        : /mono|courier|consol/.test(lower)\n          ? 'monospace'\n          : 'sans-serif'\n    return `${style} ${weight} ${Math.max(1, Math.round(size))}px ${family}`\n}\n\n/**\n * Relative advance width per character, used when no canvas is available.\n * Buckets rather than real metrics -- enough to keep proportions sane.\n */\nexport function relativeCharWidth(char: string): number {\n    if (\"iljI|.,:;'`!\".includes(char)) return 0.6\n    if ('ftr()[]{}-'.includes(char)) return 0.8\n    if ('MWmw@%'.includes(char)) return 1.6\n    if (char === ' ') return 0.6\n    if (char >= 'A' && char <= 'Z') return 1.3\n    return 1.0\n}\n\nfunction measureWord(word: string, font: string): number {\n    const ctx = measurementContext()\n    if (ctx) {\n        ctx.font = font\n        return ctx.measureText(word).width\n    }\n    let total = 0\n    for (const char of word) total += relativeCharWidth(char)\n    return total\n}\n\n/**\n * Split one run into word boxes.\n *\n * Returns the run itself when it holds a single word, so the common case costs\n * nothing.\n */\nexport function splitRunIntoWords(run: TextBox): Box[] {\n    const trimmed = run.text.trim()\n    if (trimmed.length === 0) return []\n\n    const tokens = trimmed.split(/(\\s+)/).filter((t) => t.length > 0)\n    const words = tokens.filter((t) => !/^\\s+$/.test(t))\n    if (words.length <= 1) {\n        return [{ ...run, text: trimmed }]\n    }\n\n    const font = cssFontFromPdfName(run.fontName, run.height)\n    const measured = tokens.map((token) => measureWord(token, font))\n    const totalMeasured = measured.reduce((sum, w) => sum + w, 0) || 1\n\n    // The substitute font's absolute metrics are meaningless; only the ratios\n    // matter, so normalise them onto the run's real extent.\n    const runLength = Math.hypot(run.width * run.readDirX, run.height * run.readDirY) || run.width\n    const scale = runLength / totalMeasured\n\n    // Walking starts at whichever corner the reading direction comes *from*, so\n    // a right-to-left or bottom-to-top run starts at the opposite edge.\n    let cursorX = run.readDirX >= 0 ? run.x : run.x + run.width\n    let cursorY = run.readDirY >= 0 ? run.y : run.y + run.height\n\n    const out: Box[] = []\n    for (const [i, token] of tokens.entries()) {\n        const advance = (measured[i] as number) * scale\n        if (!/^\\s+$/.test(token)) {\n            const pad = run.height * WORD_PADDING_RATIO\n            const spanX = Math.abs(run.readDirX) > Math.abs(run.readDirY) ? advance : run.width\n            const spanY = Math.abs(run.readDirY) > Math.abs(run.readDirX) ? advance : run.height\n\n            const left = run.readDirX >= 0 ? cursorX : cursorX - spanX\n            const top = run.readDirY >= 0 ? cursorY : cursorY - spanY\n\n            out.push({\n                text: token,\n                score: run.score,\n                x: Math.floor(left - pad),\n                y: Math.floor(top - pad),\n                width: Math.max(1, Math.ceil(spanX + pad * 2)),\n                height: Math.max(1, Math.ceil(spanY + pad * 2)),\n                angle: run.angle,\n            })\n        }\n        cursorX += run.readDirX * advance\n        cursorY += run.readDirY * advance\n    }\n    return out\n}\n\n/** Split every run on a page into word boxes. */\nexport function splitRunsIntoWords(runs: readonly TextBox[]): Box[] {\n    return runs.flatMap(splitRunIntoWords)\n}\n\n/** Reset the cached measurement canvas. Tests only. */\nexport function resetMeasurementContext(): void {\n    measureCtx = null\n}\n","/**\n * Port of `scaledp/pdf/PdfDataToText.py`: a PDF's embedded text layer into one\n * `Document` row per page, with word-level boxes.\n *\n * Coordinates are emitted in the same pixel space `PdfToImage` renders at, so\n * boxes from this stage and boxes from OCR are directly comparable. Python\n * leaves PdfDataToText in PDF points and scales only in PdfDataToDocument; a\n * single consistent space is more useful and avoids a class of silent mismatch.\n *\n * The output feeds the `bypassCol` optimisation: a page that already has a text\n * layer does not need OCR.\n */\n\nimport { ImageError } from '../core/errors.js'\nimport { BASE_STAGE_DEFAULTS, type BaseStageParams, resolveParams } from '../core/params.js'\nimport { type Row, Stage, type StageContext } from '../core/pipeline.js'\nimport { createDocument, type Document } from '../schemas/document.js'\nimport { toBytes } from '../stages/data-to-image.js'\nimport { extractTextBoxes } from './extract-text.js'\nimport { POINTS_PER_INCH } from './pdf-to-image.js'\nimport { describePdfError, documentOptions, loadPdfjs } from './pdfjs.js'\nimport { splitRunsIntoWords } from './split-words.js'\n\nexport interface PdfToDocumentParams extends BaseStageParams {\n    /** Pixel space the boxes are expressed in; match PdfToImage to align them. */\n    resolution: number\n    pageLimit: number\n    /** Split pdf.js line runs into word boxes. Off yields run-level boxes. */\n    splitWords: boolean\n}\n\nexport const PDF_TO_DOCUMENT_DEFAULTS: PdfToDocumentParams = Object.freeze({\n    ...BASE_STAGE_DEFAULTS,\n    inputCol: 'content',\n    outputCol: 'document',\n    keepInputData: true,\n    resolution: 300,\n    pageLimit: 0,\n    splitWords: true,\n})\n\nexport class PdfToDocument extends Stage<PdfToDocumentParams> {\n    readonly name = 'PdfToDocument'\n\n    constructor(options: Partial<PdfToDocumentParams> = {}) {\n        super(resolveParams(PDF_TO_DOCUMENT_DEFAULTS, options))\n    }\n\n    protected override async expand(input: unknown, row: Row, ctx: StageContext): Promise<Row[]> {\n        const { outputCol, pageCol, pathCol, resolution, pageLimit, splitWords } = this.params\n        const path = String(row[pathCol] ?? 'memory')\n\n        const pdfjs = await loadPdfjs()\n        const task = pdfjs.getDocument(documentOptions(toBytes(input)))\n\n        try {\n            // pdf.js defers worker setup, so a missing worker surfaces on first\n            // page access rather than from task.promise.\n            const pdf = await task.promise\n            const pageCount = pageLimit > 0 ? Math.min(pageLimit, pdf.numPages) : pdf.numPages\n            const rows: Row[] = []\n\n            for (let index = 0; index < pageCount; index++) {\n                ctx.signal?.throwIfAborted()\n                const page = await pdf.getPage(index + 1)\n                try {\n                    const viewport = page.getViewport({ scale: resolution / POINTS_PER_INCH })\n                    const runs = await extractTextBoxes(page, viewport)\n                    const bboxes = splitWords ? splitRunsIntoWords(runs) : runs\n\n                    rows.push({\n                        ...row,\n                        [pageCol]: index,\n                        [outputCol]: createDocument({\n                            path,\n                            type: 'pdf',\n                            text: runs.map((r) => r.text).join('\\n'),\n                            bboxes,\n                        }),\n                    })\n                } finally {\n                    page.cleanup()\n                }\n            }\n            return rows\n        } catch (error) {\n            throw describePdfError(error)\n        } finally {\n            await task.destroy()\n        }\n    }\n\n    protected async apply(): Promise<never> {\n        throw new ImageError('unreachable: expand handles every row', this.name)\n    }\n\n    protected onError(message: string, row: Row): Document {\n        return createDocument({\n            path: String(row[this.params.pathCol] ?? 'memory'),\n            type: 'pdf',\n            exception: message,\n        })\n    }\n}\n\n/** True when a page's text layer is substantive enough to skip OCR. */\nexport function hasUsableTextLayer(document: Document, minimumBoxes = 1): boolean {\n    return document.exception === '' && document.bboxes.length >= minimumBoxes\n}\n"],"mappings":";;;;;;;AA+BA,MAAM,eAAe;;AAGrB,MAAa,mBAAmB;AAchC,SAAgB,WAAW,MAAqC;CAC5D,OACI,OAAO,SAAS,YAChB,SAAS,QACT,SAAS,QACT,eAAe,QACf,MAAM,QAAS,KAAsB,SAAS;AAEtD;;AAGA,SAAgB,cAAc,MAAoB,UAAwC;CACtF,IAAI,KAAK,IAAI,WAAW,GAAG,OAAO;CAElC,MAAM,CAAC,IAAI,GAAG,IAAI,GAAG,IAAI,GAAG,IAAI,GAAG,IAAI,GAAG,IAAI,KAAK,KAAK;CAIxD,MAAM,QAAQ,KAAK,MAAM,GAAG,CAAC,KAAK;CAClC,MAAM,QAAQ,KAAK,MAAM,GAAG,CAAC,KAAK;CAClC,MAAM,OAAO,IAAI;CACjB,MAAM,OAAO,IAAI;CACjB,MAAM,MAAM,IAAI;CAChB,MAAM,MAAM,IAAI;CAGhB,MAAM,SAAS,KAAK,SAAS;CAC7B,MAAM,SAAS,IAAI,MAAM;CACzB,MAAM,SAAS,IAAI,MAAM;CAEzB,MAAM,UAAU;EACZ,SAAS,uBAAuB,QAAQ,MAAM;EAC9C,SAAS,uBAAuB,SAAS,OAAO,KAAK,OAAO,SAAS,OAAO,KAAK,KAAK;EACtF,SAAS,uBAAuB,SAAS,MAAM,KAAK,QAAQ,SAAS,MAAM,KAAK,MAAM;EACtF,SAAS,uBACL,SAAS,OAAO,KAAK,QAAQ,MAAM,KAAK,QACxC,SAAS,OAAO,KAAK,QAAQ,MAAM,KAAK,MAC5C;CACJ;CAEA,MAAM,KAAK,QAAQ,KAAK,MAAM,EAAE,EAAY;CAC5C,MAAM,KAAK,QAAQ,KAAK,MAAM,EAAE,EAAY;CAC5C,MAAM,IAAI,KAAK,IAAI,GAAG,EAAE;CACxB,MAAM,IAAI,KAAK,IAAI,GAAG,EAAE;CAIxB,MAAM,CAAC,eAAe,GAAG,eAAe,KAAK,QAAQ,MAAM,CAAC;CAC5D,MAAM,CAAC,aAAa,GAAG,aAAa,KAAK,QAAQ,MAAM,CAAC;CACxD,MAAM,UAAU,KAAK,MAAM,aAAa,cAAc,aAAa,YAAY,KAAK;CAEpF,OAAO;EACH,MAAM,KAAK;EACX,OAAO;EACP,GAAG,KAAK,MAAM,CAAC;EACf,GAAG,KAAK,MAAM,CAAC;EACf,OAAO,KAAK,IAAI,GAAG,KAAK,KAAK,KAAK,IAAI,GAAG,EAAE,IAAI,CAAC,CAAC;EACjD,QAAQ,KAAK,IAAI,GAAG,KAAK,KAAK,KAAK,IAAI,GAAG,EAAE,IAAI,CAAC,CAAC;EAClD,OAAO;EACP,WAAW,aAAa,gBAAgB;EACxC,WAAW,aAAa,gBAAgB;EACxC,UAAU,KAAK,YAAY;CAC/B;AACJ;;AAGA,eAAsB,iBAClB,MACA,UACkB;CAClB,MAAM,UAAU,MAAM,KAAK,eAAe;CAC1C,MAAM,QAAmB,CAAC;CAC1B,KAAK,MAAM,QAAQ,QAAQ,OAAO;EAC9B,IAAI,CAAC,WAAW,IAAI,GAAG;EACvB,MAAM,MAAM,cAAc,MAAM,QAAQ;EACxC,IAAI,KAAK,MAAM,KAAK,GAAG;CAC3B;CACA,OAAO;AACX;;;;;;;;;;;ACjHA,IAAI,gBAA6C;AAEjD,eAAsB,YAAkC;CACpD,IAAI,eAAe,OAAO;CAE1B,iBAAiB,YAAY;EACzB,IAAI;EACJ,IAAI;GACA,QAAQ,MAAM,OAAO;EACzB,SAAS,OAAO;GACZ,MAAM,IAAI,MAAM,wEAAwE,EAAE,MAAM,CAAC;EACrG;EAEA,MAAM,EAAE,cAAc,UAAU,CAAC,CAAC;EAClC,IAAI,WAAW,MAAM,oBAAoB,YAAY;EACrD,OAAO;CACX,EAAA,CAAG;CAEH,OAAO;AACX;;AAGA,SAAgB,aAAmB;CAC/B,gBAAgB;AACpB;;;;;;;;;AAUA,SAAgB,iBAAiB,OAAuB;CACpD,MAAM,UAAU,iBAAiB,QAAQ,MAAM,UAAU,OAAO,KAAK;CACrE,IAAI,CAAC,sBAAsB,KAAK,OAAO,GACnC,OAAO,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO;CAG7D,MAAM,EAAE,cAAc,UAAU,CAAC,CAAC;CAClC,MAAM,QAAQ,YACR,0CAA0C,UAAU,MACpD;CAEN,OAAO,IAAI,MACP,GAAG,MAAM;;;kBAIc,WACvB,EAAE,OAAO,MAAM,CACnB;AACJ;;AAGA,SAAgB,gBAAgB,MAA2C;CACvE,MAAM,EAAE,SAAS,qBAAqB,YAAY,UAAU,CAAC,CAAC;CAG9D,MAAM,QAAQ,IAAI,WAAW,KAAK,UAAU;CAC5C,MAAM,IAAI,IAAI;CAEd,MAAM,UAAmC,EAAE,MAAM,MAAM;CACvD,IAAI,SAAS;EACT,QAAQ,UAAU;EAClB,QAAQ,aAAa;CACzB;CACA,IAAI,qBAAqB,QAAQ,sBAAsB;CACvD,IAAI,SAAS,QAAQ,UAAU;CAC/B,OAAO;AACX;;;;;;;;;;ACpEA,MAAa,kBAAkB;AAU/B,MAAa,wBAA0C,OAAO,OAAO;CACjE,GAAG;CACH,UAAU;CACV,WAAW;CACX,YAAY;CACZ,WAAW;CACX,WAAW;AACf,CAAC;AAED,MAAM,OAAuE;CACzE,KAAK;CACL,MAAM;CACN,MAAM;AACV;AAEA,IAAa,aAAb,cAAgC,MAAwB;CACpD,OAAgB;CAEhB,YAAY,UAAqC,CAAC,GAAG;EACjD,MACI,cAAc,uBAAuB,SAAS;GAC1C,aAAa,UAAU;IACnB,IAAI,CAAC,OAAO,SAAS,KAAK,KAAK,SAAS,GACpC,MAAM,IAAI,WAAW,yCAAyC,OAAO;GAE7E;GACA,YAAY,UAAU;IAClB,IAAI,CAAC,OAAO,UAAU,KAAK,KAAK,QAAQ,GACpC,MAAM,IAAI,WAAW,sDAAsD,OAAO;GAE1F;EACJ,CAAC,CACL;CACJ;;CAGA,MAAyB,OAAO,OAAgB,KAAU,KAAmC;EACzF,MAAM,EAAE,WAAW,SAAS,SAAS,YAAY,WAAW,cAAc,KAAK;EAC/E,MAAM,OAAO,OAAO,IAAI,YAAY,QAAQ;EAG5C,MAAM,QAAO,MADO,UAAU,EAAA,CACX,YAAY,gBAAgB,QAAQ,KAAK,CAAC,CAAC;EAE9D,IAAI;GAIA,MAAM,WAAW,MAAM,KAAK;GAC5B,MAAM,YAAY,YAAY,IAAI,KAAK,IAAI,WAAW,SAAS,QAAQ,IAAI,SAAS;GACpF,MAAM,OAAc,CAAC;GAErB,KAAK,IAAI,QAAQ,GAAG,QAAQ,WAAW,SAAS;IAC5C,IAAI,QAAQ,eAAe;IAC3B,MAAM,QAAQ,MAAM,WAAW,UAAU,QAAQ,GAAG;KAChD;KACA;KACA;IACJ,CAAC;IACD,KAAK,KAAK;KAAE,GAAG;MAAM,UAAU;MAAQ,YAAY;IAAM,CAAC;GAC9D;GACA,OAAO;EACX,SAAS,OAAO;GACZ,MAAM,iBAAiB,KAAK;EAChC,UAAU;GAGN,MAAM,KAAK,QAAQ;EACvB;CACJ;CAEA,MAAgB,QAAwB;EACpC,MAAM,IAAI,WAAW,yCAAyC,KAAK,IAAI;CAC3E;CAEA,QAAkB,SAAiB,KAAwB;EACvD,OAAO,YAAY;GAAE,MAAM,OAAO,IAAI,KAAK,OAAO,YAAY,QAAQ;GAAG,WAAW;EAAQ,CAAC;CACjG;AACJ;;AAGA,eAAsB,WAClB,UACA,YACA,MACqB;CACrB,MAAM,OAAO,MAAM,SAAS,QAAQ,UAAU;CAC9C,IAAI;EACA,MAAM,WAAW,KAAK,YAAY,EAAE,OAAO,KAAK,aAAA,GAA6B,CAAC;EAC9E,MAAM,SAAS,aAAa,SAAS,OAAO,SAAS,MAAM;EAI3D,MAAM,KAAK,OAAO;GACN;GACR;EACJ,CAAC,CAAC,CAAC;EAEH,OAAO,YAAY;GACf,MAAM,KAAK;GACX,YAAY,KAAK;GACjB,MAAM,MAAM,YAAY,QAAQ,KAAK,KAAK,UAAU;GACpD,WAAW,KAAK;GAChB,OAAO,OAAO;GACd,QAAQ,OAAO;EACnB,CAAC;CACL,UAAU;EACN,KAAK,QAAQ;CACjB;AACJ;;;;;;;;;;;;;;;;;ACnHA,MAAM,qBAAqB;AAE3B,IAAI,aAAuD;AAE3D,SAAS,qBAA+D;CACpE,IAAI,YAAY,OAAO;CACvB,IAAI;EACA,aAAa,UAAU,aAAa,GAAG,CAAC,CAAC;EACzC,OAAO;CACX,QAAQ;EAEJ,OAAO;CACX;AACJ;;;;;;;AAQA,SAAgB,mBAAmB,UAAkB,MAAsB;CAEvE,MAAM,QADO,SAAS,QAAQ,eAAe,EAC5B,CAAC,CAAC,YAAY;CAC/B,MAAM,SAAS,4BAA4B,KAAK,KAAK,IAAI,SAAS;CAClE,MAAM,QAAQ,iBAAiB,KAAK,KAAK,IAAI,WAAW;CACxD,MAAM,SAAS,qCAAqC,KAAK,KAAK,IACxD,UACA,sBAAsB,KAAK,KAAK,IAC9B,cACA;CACR,OAAO,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,IAAI,GAAG,KAAK,MAAM,IAAI,CAAC,EAAE,KAAK;AACpE;;;;;AAMA,SAAgB,kBAAkB,MAAsB;CACpD,IAAI,eAAe,SAAS,IAAI,GAAG,OAAO;CAC1C,IAAI,aAAa,SAAS,IAAI,GAAG,OAAO;CACxC,IAAI,SAAS,SAAS,IAAI,GAAG,OAAO;CACpC,IAAI,SAAS,KAAK,OAAO;CACzB,IAAI,QAAQ,OAAO,QAAQ,KAAK,OAAO;CACvC,OAAO;AACX;AAEA,SAAS,YAAY,MAAc,MAAsB;CACrD,MAAM,MAAM,mBAAmB;CAC/B,IAAI,KAAK;EACL,IAAI,OAAO;EACX,OAAO,IAAI,YAAY,IAAI,CAAC,CAAC;CACjC;CACA,IAAI,QAAQ;CACZ,KAAK,MAAM,QAAQ,MAAM,SAAS,kBAAkB,IAAI;CACxD,OAAO;AACX;;;;;;;AAQA,SAAgB,kBAAkB,KAAqB;CACnD,MAAM,UAAU,IAAI,KAAK,KAAK;CAC9B,IAAI,QAAQ,WAAW,GAAG,OAAO,CAAC;CAElC,MAAM,SAAS,QAAQ,MAAM,OAAO,CAAC,CAAC,QAAQ,MAAM,EAAE,SAAS,CAAC;CAEhE,IADc,OAAO,QAAQ,MAAM,CAAC,QAAQ,KAAK,CAAC,CAC1C,CAAC,CAAC,UAAU,GAChB,OAAO,CAAC;EAAE,GAAG;EAAK,MAAM;CAAQ,CAAC;CAGrC,MAAM,OAAO,mBAAmB,IAAI,UAAU,IAAI,MAAM;CACxD,MAAM,WAAW,OAAO,KAAK,UAAU,YAAY,OAAO,IAAI,CAAC;CAC/D,MAAM,gBAAgB,SAAS,QAAQ,KAAK,MAAM,MAAM,GAAG,CAAC,KAAK;CAKjE,MAAM,SADY,KAAK,MAAM,IAAI,QAAQ,IAAI,UAAU,IAAI,SAAS,IAAI,QAAQ,KAAK,IAAI,SAC/D;CAI1B,IAAI,UAAU,IAAI,YAAY,IAAI,IAAI,IAAI,IAAI,IAAI,IAAI;CACtD,IAAI,UAAU,IAAI,YAAY,IAAI,IAAI,IAAI,IAAI,IAAI,IAAI;CAEtD,MAAM,MAAa,CAAC;CACpB,KAAK,MAAM,CAAC,GAAG,UAAU,OAAO,QAAQ,GAAG;EACvC,MAAM,UAAW,SAAS,KAAgB;EAC1C,IAAI,CAAC,QAAQ,KAAK,KAAK,GAAG;GACtB,MAAM,MAAM,IAAI,SAAS;GACzB,MAAM,QAAQ,KAAK,IAAI,IAAI,QAAQ,IAAI,KAAK,IAAI,IAAI,QAAQ,IAAI,UAAU,IAAI;GAC9E,MAAM,QAAQ,KAAK,IAAI,IAAI,QAAQ,IAAI,KAAK,IAAI,IAAI,QAAQ,IAAI,UAAU,IAAI;GAE9E,MAAM,OAAO,IAAI,YAAY,IAAI,UAAU,UAAU;GACrD,MAAM,MAAM,IAAI,YAAY,IAAI,UAAU,UAAU;GAEpD,IAAI,KAAK;IACL,MAAM;IACN,OAAO,IAAI;IACX,GAAG,KAAK,MAAM,OAAO,GAAG;IACxB,GAAG,KAAK,MAAM,MAAM,GAAG;IACvB,OAAO,KAAK,IAAI,GAAG,KAAK,KAAK,QAAQ,MAAM,CAAC,CAAC;IAC7C,QAAQ,KAAK,IAAI,GAAG,KAAK,KAAK,QAAQ,MAAM,CAAC,CAAC;IAC9C,OAAO,IAAI;GACf,CAAC;EACL;EACA,WAAW,IAAI,WAAW;EAC1B,WAAW,IAAI,WAAW;CAC9B;CACA,OAAO;AACX;;AAGA,SAAgB,mBAAmB,MAAiC;CAChE,OAAO,KAAK,QAAQ,iBAAiB;AACzC;;AAGA,SAAgB,0BAAgC;CAC5C,aAAa;AACjB;;;;;;;;;;;;;;;AC/GA,MAAa,2BAAgD,OAAO,OAAO;CACvE,GAAG;CACH,UAAU;CACV,WAAW;CACX,eAAe;CACf,YAAY;CACZ,WAAW;CACX,YAAY;AAChB,CAAC;AAED,IAAa,gBAAb,cAAmC,MAA2B;CAC1D,OAAgB;CAEhB,YAAY,UAAwC,CAAC,GAAG;EACpD,MAAM,cAAc,0BAA0B,OAAO,CAAC;CAC1D;CAEA,MAAyB,OAAO,OAAgB,KAAU,KAAmC;EACzF,MAAM,EAAE,WAAW,SAAS,SAAS,YAAY,WAAW,eAAe,KAAK;EAChF,MAAM,OAAO,OAAO,IAAI,YAAY,QAAQ;EAG5C,MAAM,QAAO,MADO,UAAU,EAAA,CACX,YAAY,gBAAgB,QAAQ,KAAK,CAAC,CAAC;EAE9D,IAAI;GAGA,MAAM,MAAM,MAAM,KAAK;GACvB,MAAM,YAAY,YAAY,IAAI,KAAK,IAAI,WAAW,IAAI,QAAQ,IAAI,IAAI;GAC1E,MAAM,OAAc,CAAC;GAErB,KAAK,IAAI,QAAQ,GAAG,QAAQ,WAAW,SAAS;IAC5C,IAAI,QAAQ,eAAe;IAC3B,MAAM,OAAO,MAAM,IAAI,QAAQ,QAAQ,CAAC;IACxC,IAAI;KAEA,MAAM,OAAO,MAAM,iBAAiB,MADnB,KAAK,YAAY,EAAE,OAAO,aAAA,GAA6B,CAC9B,CAAQ;KAClD,MAAM,SAAS,aAAa,mBAAmB,IAAI,IAAI;KAEvD,KAAK,KAAK;MACN,GAAG;OACF,UAAU;OACV,YAAY,eAAe;OACxB;OACA,MAAM;OACN,MAAM,KAAK,KAAK,MAAM,EAAE,IAAI,CAAC,CAAC,KAAK,IAAI;OACvC;MACJ,CAAC;KACL,CAAC;IACL,UAAU;KACN,KAAK,QAAQ;IACjB;GACJ;GACA,OAAO;EACX,SAAS,OAAO;GACZ,MAAM,iBAAiB,KAAK;EAChC,UAAU;GACN,MAAM,KAAK,QAAQ;EACvB;CACJ;CAEA,MAAgB,QAAwB;EACpC,MAAM,IAAI,WAAW,yCAAyC,KAAK,IAAI;CAC3E;CAEA,QAAkB,SAAiB,KAAoB;EACnD,OAAO,eAAe;GAClB,MAAM,OAAO,IAAI,KAAK,OAAO,YAAY,QAAQ;GACjD,MAAM;GACN,WAAW;EACf,CAAC;CACL;AACJ;;AAGA,SAAgB,mBAAmB,UAAoB,eAAe,GAAY;CAC9E,OAAO,SAAS,cAAc,MAAM,SAAS,OAAO,UAAU;AAClE"}