@datasynx/agentic-crm 1.0.0 → 1.1.0

package/dist/attachments-rLa96rOK.js

@@ -0,0 +1,514 @@
+import { t as assertSafeSlug } from "./customer-dir-CkMMXhb0.js";
+import { n as logger } from "./logger-Dyl4VcLO.js";
+import { t as chunkText } from "./chunk-BhUZmQg5.js";
+import { n as htmlToMarkdown, t as htmlConverter } from "./html-BaeOCZKE.js";
+import path from "path";
+import fs from "fs";
+//#region src/sync/converters/text.ts
+/** Escape a CSV cell for safe inclusion in a Markdown table cell. */
+function mdCell(value) {
+	return value.replace(/\\/g, "\\\\").replace(/\|/g, "\\|").replace(/\r?\n/g, " ").trim();
+}
+/**
+* Minimal RFC-4180-ish CSV line splitter: handles quoted fields containing
+* commas and escaped double quotes. Good enough for rendering CSV attachments
+* as readable Markdown tables (we are not round-tripping data).
+*/
+function parseCsvLine(line) {
+	const out = [];
+	let cur = "";
+	let inQuotes = false;
+	for (let i = 0; i < line.length; i++) {
+		const ch = line[i];
+		if (inQuotes) if (ch === "\"") if (line[i + 1] === "\"") {
+			cur += "\"";
+			i++;
+		} else inQuotes = false;
+		else cur += ch;
+		else if (ch === "\"") inQuotes = true;
+		else if (ch === ",") {
+			out.push(cur);
+			cur = "";
+		} else cur += ch;
+	}
+	out.push(cur);
+	return out;
+}
+/** Render a matrix of cells as a GitHub-flavored Markdown pipe table. */
+function rowsToMarkdown(rows) {
+	if (rows.length === 0) return "";
+	const width = Math.max(...rows.map((r) => r.length));
+	const pad = (r) => Array.from({ length: width }, (_, i) => mdCell(r[i] ?? ""));
+	const header = pad(rows[0] ?? []);
+	return [
+		`| ${header.join(" | ")} |`,
+		`| ${header.map(() => "---").join(" | ")} |`,
+		...rows.slice(1).map((r) => `| ${pad(r).join(" | ")} |`)
+	].join("\n");
+}
+/** Render CSV text as a GitHub-flavored Markdown pipe table. */
+function csvToMarkdown(csv) {
+	return rowsToMarkdown(csv.split(/\r?\n/).filter((l) => l.trim().length > 0).map(parseCsvLine));
+}
+const TEXT_EXTENSIONS = [
+	"txt",
+	"text",
+	"log",
+	"md",
+	"markdown"
+];
+const CODE_FENCE_EXTENSIONS = {
+	json: "json",
+	xml: "xml",
+	yaml: "yaml",
+	yml: "yaml"
+};
+/**
+* Converter for plain-text-ish attachments: Markdown/text passthrough, CSV/TSV
+* to Markdown tables, and structured text (JSON/XML/YAML) into fenced code
+* blocks so they stay readable and searchable without a heavy parser.
+*/
+const textConverter = {
+	name: "text",
+	extensions: [
+		...TEXT_EXTENSIONS,
+		"csv",
+		"tsv",
+		...Object.keys(CODE_FENCE_EXTENSIONS)
+	],
+	mimeTypes: [
+		"text/plain",
+		"text/csv",
+		"text/markdown",
+		"application/json",
+		"text/*"
+	],
+	convert(buffer, filename) {
+		const ext = filename.split(".").pop()?.toLowerCase() ?? "";
+		const content = buffer.toString("utf-8");
+		if (ext === "csv") return Promise.resolve({
+			markdown: csvToMarkdown(content),
+			meta: { format: "csv" }
+		});
+		if (ext === "tsv") {
+			const asCsv = content.replace(/\t/g, ",");
+			return Promise.resolve({
+				markdown: csvToMarkdown(asCsv),
+				meta: { format: "tsv" }
+			});
+		}
+		const fence = CODE_FENCE_EXTENSIONS[ext];
+		if (fence) return Promise.resolve({
+			markdown: `\`\`\`${fence}\n${content.trim()}\n\`\`\``,
+			meta: { format: fence }
+		});
+		return Promise.resolve({
+			markdown: content.trim(),
+			meta: { format: "text" }
+		});
+	}
+};
+//#endregion
+//#region src/sync/converters/docx.ts
+/**
+* DOCX → Markdown via mammoth (DOCX → semantic HTML) then Turndown (HTML →
+* Markdown). Mammoth's own Markdown output is deprecated; the HTML route keeps
+* tables, lists and headings intact. mammoth is loaded lazily.
+*/
+const docxConverter = {
+	name: "docx",
+	extensions: ["docx"],
+	mimeTypes: ["application/vnd.openxmlformats-officedocument.wordprocessingml.document"],
+	async convert(buffer) {
+		const { value: html, messages } = await (await import("mammoth")).default.convertToHtml({ buffer });
+		return {
+			markdown: await htmlToMarkdown(html),
+			meta: {
+				format: "docx",
+				warnings: messages.filter((m) => m.type === "warning").length
+			}
+		};
+	}
+};
+//#endregion
+//#region src/sync/converters/xlsx.ts
+/** Render a single ExcelJS cell value as plain text. */
+function cellText(value) {
+	if (value === null || value === void 0) return "";
+	if (value instanceof Date) return value.toISOString().slice(0, 10);
+	if (typeof value === "object") {
+		const v = value;
+		if (typeof v["text"] === "string") return v["text"];
+		if ("result" in v) return String(v["result"] ?? "");
+		if (Array.isArray(v["richText"])) return v["richText"].map((r) => r.text ?? "").join("");
+		if ("hyperlink" in v) return String(v["text"] ?? v["hyperlink"] ?? "");
+	}
+	return String(value);
+}
+/**
+* Spreadsheet (XLSX) → Markdown via ExcelJS. Each worksheet becomes a
+* `## <sheet name>` section followed by a GitHub-flavored Markdown table.
+* ExcelJS is loaded lazily.
+*/
+const xlsxConverter = {
+	name: "xlsx",
+	extensions: ["xlsx", "xlsm"],
+	mimeTypes: ["application/vnd.openxmlformats-officedocument.spreadsheetml.sheet", "application/vnd.ms-excel"],
+	async convert(buffer) {
+		const wb = new (await (import("exceljs"))).default.Workbook();
+		await wb.xlsx.load(buffer);
+		const sections = [];
+		const sheetNames = [];
+		wb.eachSheet((sheet) => {
+			sheetNames.push(sheet.name);
+			const rows = [];
+			sheet.eachRow({ includeEmpty: false }, (row) => {
+				const cells = [];
+				row.eachCell({ includeEmpty: true }, (cell) => {
+					cells.push(cellText(cell.value));
+				});
+				rows.push(cells);
+			});
+			const table = rowsToMarkdown(rows);
+			if (table) sections.push(`## ${sheet.name}\n\n${table}`);
+		});
+		return {
+			markdown: sections.join("\n\n"),
+			meta: {
+				format: "xlsx",
+				sheets: sheetNames
+			}
+		};
+	}
+};
+//#endregion
+//#region src/sync/converters/pptx.ts
+/** Extract the visible text runs (`<a:t>…</a:t>`) from one slide's XML. */
+function extractSlideText(xml) {
+	return [...xml.matchAll(/<a:t>([\s\S]*?)<\/a:t>/g)].map((m) => (m[1] ?? "").replace(/&amp;/g, "&").replace(/&lt;/g, "<").replace(/&gt;/g, ">").replace(/&quot;/g, "\"").replace(/&apos;/g, "'")).join(" ").replace(/\s+/g, " ").trim();
+}
+function slideNumber(entryName) {
+	const m = entryName.match(/slide(\d+)\.xml$/);
+	return m ? parseInt(m[1] ?? "0", 10) : 0;
+}
+/**
+* PPTX → Markdown. A .pptx is a zip; slide text lives in `ppt/slides/slideN.xml`
+* as `<a:t>` runs. We unzip with adm-zip (already a dependency) and emit one
+* `## Slide N` section per slide — no extra native parser needed.
+*/
+const pptxConverter = {
+	name: "pptx",
+	extensions: ["pptx"],
+	mimeTypes: ["application/vnd.openxmlformats-officedocument.presentationml.presentation"],
+	async convert(buffer) {
+		const AdmZip = (await import("adm-zip")).default;
+		const slides = new AdmZip(buffer).getEntries().filter((e) => /^ppt\/slides\/slide\d+\.xml$/.test(e.entryName)).sort((a, b) => slideNumber(a.entryName) - slideNumber(b.entryName));
+		const sections = [];
+		for (const entry of slides) {
+			const text = extractSlideText(entry.getData().toString("utf-8"));
+			if (text) sections.push(`## Slide ${slideNumber(entry.entryName)}\n\n${text}`);
+		}
+		return {
+			markdown: sections.join("\n\n"),
+			meta: {
+				format: "pptx",
+				slides: slides.length
+			}
+		};
+	}
+};
+//#endregion
+//#region src/sync/converters/pdf.ts
+/** Whether scanned-PDF OCR is enabled (opt-in, requires @napi-rs/canvas). */
+function isPdfOcrEnabled() {
+	const v = (process.env["DXCRM_PDF_OCR"] ?? "").toLowerCase();
+	return v === "1" || v === "true" || v === "yes" || v === "on";
+}
+/**
+* OCR a scanned PDF by rendering each page to a PNG (unpdf + @napi-rs/canvas)
+* and running tesseract.js over it. The canvas backend is an optional peer
+* dependency; if it isn't installed we throw a clear, catchable error so the
+* caller can fall back to the OCR-candidate stub. Page count is capped via
+* DXCRM_PDF_OCR_MAX_PAGES (default 20) to keep this slow path bounded.
+*/
+async function ocrPdf(pdf, totalPages) {
+	try {
+		await import("@napi-rs/canvas");
+	} catch {
+		throw new Error("PDF OCR requires the optional '@napi-rs/canvas' package — run `npm install @napi-rs/canvas`");
+	}
+	const { renderPageAsImage } = await import("unpdf");
+	const { recognize } = await import("tesseract.js");
+	const lang = process.env["DXCRM_OCR_LANG"] ?? "eng";
+	const canvasImport = () => import("@napi-rs/canvas");
+	const maxPages = Number(process.env["DXCRM_PDF_OCR_MAX_PAGES"] ?? 20) || 20;
+	const pageCount = Math.min(totalPages, maxPages);
+	const pages = [];
+	for (let p = 1; p <= pageCount; p++) {
+		const png = await renderPageAsImage(pdf, p, {
+			canvasImport,
+			scale: 2
+		});
+		const { data: { text } } = await recognize(Buffer.from(png), lang);
+		if (text.trim()) pages.push(`## Page ${p}\n\n${text.trim()}`);
+	}
+	return pages.join("\n\n");
+}
+/**
+* PDF → Markdown. Extracts the digital text layer with unpdf (a serverless
+* pdf.js build). Scanned PDFs have no text layer: when DXCRM_PDF_OCR is enabled
+* they are rendered and OCR'd page-by-page, otherwise they're flagged as OCR
+* candidates rather than emitting garbage. unpdf is loaded lazily.
+*/
+const pdfConverter = {
+	name: "pdf",
+	extensions: ["pdf"],
+	mimeTypes: ["application/pdf"],
+	async convert(buffer, filename) {
+		const { extractText, getDocumentProxy } = await import("unpdf");
+		const pdf = await getDocumentProxy(new Uint8Array(buffer));
+		const { totalPages, text } = await extractText(pdf, { mergePages: true });
+		const merged = (Array.isArray(text) ? text.join("\n\n") : text).trim();
+		if (merged) return {
+			markdown: merged,
+			meta: {
+				format: "pdf",
+				pages: totalPages
+			}
+		};
+		if (isPdfOcrEnabled()) try {
+			const ocr = await ocrPdf(pdf, totalPages);
+			if (ocr.trim()) return {
+				markdown: `> _OCR of \`${filename}\` (${totalPages} pages):_\n\n${ocr.trim()}`,
+				meta: {
+					format: "pdf",
+					pages: totalPages,
+					ocr: true
+				}
+			};
+		} catch (err) {
+			process.stderr.write(`[converters] pdf OCR failed for ${filename}: ${err.message}\n`);
+		}
+		return {
+			markdown: "",
+			meta: {
+				format: "pdf",
+				pages: totalPages,
+				ocrCandidate: true
+			}
+		};
+	}
+};
+//#endregion
+//#region src/sync/converters/image.ts
+/**
+* Image → Markdown via Tesseract.js OCR (pure-JS, 100+ languages, fully local).
+* This is the heaviest converter: tesseract.js downloads a WASM core and
+* language data on first use, so it is loaded lazily and only invoked for image
+* attachments. Language defaults to English, override with DXCRM_OCR_LANG.
+*/
+const imageConverter = {
+	name: "image",
+	extensions: [
+		"png",
+		"jpg",
+		"jpeg",
+		"tif",
+		"tiff",
+		"bmp",
+		"webp",
+		"gif",
+		"pbm"
+	],
+	mimeTypes: ["image/*"],
+	async convert(buffer, filename) {
+		const lang = process.env["DXCRM_OCR_LANG"] ?? "eng";
+		const { recognize } = await import("tesseract.js");
+		const { data: { text } } = await recognize(buffer, lang);
+		const ocr = text.trim();
+		return {
+			markdown: ocr ? `> _OCR of \`${filename}\`:_\n\n${ocr}` : "",
+			meta: {
+				format: "image",
+				ocr: true,
+				lang
+			}
+		};
+	}
+};
+//#endregion
+//#region src/sync/converters/registry.ts
+/**
+* Fallback converter for unknown/binary attachments: emit a small metadata stub
+* instead of garbage bytes, so the attachment is still recorded and linkable.
+*/
+const fallbackConverter = {
+	name: "binary",
+	extensions: [],
+	convert(buffer, filename) {
+		const kb = Math.max(1, Math.round(buffer.length / 1024));
+		return Promise.resolve({
+			markdown: `> _Binary attachment \`${filename}\` (${kb} KB) — no text representation available._`,
+			meta: {
+				format: "binary",
+				bytes: buffer.length
+			}
+		});
+	}
+};
+/**
+* Ordered converter registry. Earlier entries win on extension conflicts. The
+* text converter is intentionally last among the "real" converters so that more
+* specific formats (html, etc.) take precedence over generic text matching.
+*/
+const CONVERTERS = [
+	docxConverter,
+	xlsxConverter,
+	pptxConverter,
+	pdfConverter,
+	imageConverter,
+	htmlConverter,
+	textConverter
+];
+function extensionOf(filename) {
+	return filename.split(".").pop()?.toLowerCase() ?? "";
+}
+function mimeMatches(converter, mime) {
+	if (!converter.mimeTypes) return false;
+	const lower = mime.toLowerCase();
+	return converter.mimeTypes.some((m) => {
+		const ml = m.toLowerCase();
+		if (ml.endsWith("/*")) return lower.startsWith(ml.slice(0, -1));
+		return ml === lower;
+	});
+}
+/**
+* Pick the converter for an attachment by file extension first (most reliable
+* for Gmail attachments, which always carry a filename), then by MIME type.
+* Returns `undefined` when nothing matches.
+*/
+function matchConverter(filename, mime) {
+	const ext = extensionOf(filename);
+	if (ext) {
+		const byExt = CONVERTERS.find((c) => c.extensions.includes(ext));
+		if (byExt) return byExt;
+	}
+	if (mime) {
+		const byMime = CONVERTERS.find((c) => mimeMatches(c, mime));
+		if (byMime) return byMime;
+	}
+}
+/**
+* Convert an attachment to Markdown, dispatching to the best converter and
+* falling back to a metadata stub. Converter errors never throw: they are
+* swallowed into the fallback so a single bad attachment can't break a sync.
+*/
+async function convertAttachment(buffer, filename, mime) {
+	const converter = matchConverter(filename, mime) ?? fallbackConverter;
+	try {
+		const result = await converter.convert(buffer, filename);
+		if (!result.markdown.trim()) return {
+			markdown: `> _Attachment \`${filename}\` contained no extractable text._`,
+			meta: {
+				...result.meta,
+				empty: true
+			}
+		};
+		return result;
+	} catch (err) {
+		process.stderr.write(`[converters] ${converter.name} failed for ${filename}: ${err.message}\n`);
+		return fallbackConverter.convert(buffer, filename);
+	}
+}
+/**
+* Recursively collect downloadable attachment parts from a Gmail message
+* payload — any MIME part that carries both a filename and a body.attachmentId.
+* Inline parts without a filename (e.g. signature logos) are ignored.
+*/
+function collectAttachmentParts(payload) {
+	const out = [];
+	const walk = (part) => {
+		if (!part) return;
+		const filename = part.filename ?? "";
+		const attachmentId = part.body?.attachmentId ?? "";
+		if (filename && attachmentId) out.push({
+			filename,
+			mimeType: part.mimeType ?? "application/octet-stream",
+			attachmentId,
+			size: part.body?.size ?? 0
+		});
+		for (const child of part.parts ?? []) walk(child);
+	};
+	walk(payload);
+	return out;
+}
+/** Make a filename safe for use as a single path segment. */
+function sanitizeFilename(name) {
+	return (name.split(/[\\/]/).pop() ?? name).replace(/[^a-zA-Z0-9._-]+/g, "_").replace(/^_+|_+$/g, "").slice(0, 120) || "attachment";
+}
+/**
+* Download, convert and index every attachment of a single Gmail message.
+*
+* For each attachment: the raw bytes are saved under
+* `customers/<slug>/attachments/<messageId>__<name>`, converted to a sibling
+* `.md` file, and the Markdown is chunked and indexed into LanceDB so the
+* attachment's content is semantically searchable. Failures on a single
+* attachment are logged and skipped — they never abort the message sync.
+*/
+async function processMessageAttachments(opts) {
+	const parts = collectAttachmentParts(opts.payload);
+	if (parts.length === 0) return [];
+	assertSafeSlug(opts.slug);
+	const maxBytes = opts.maxBytes ?? 26214400;
+	const attachmentsDir = path.join(opts.dataDir, "customers", opts.slug, "attachments");
+	fs.mkdirSync(attachmentsDir, { recursive: true });
+	const { indexInLanceDB } = await import("./lancedb-CswQEE5K.js");
+	const saved = [];
+	for (const part of parts) try {
+		if (part.size > maxBytes) {
+			logger.warn("gmail-sync", "skipping oversized attachment", {
+				filename: part.filename,
+				bytes: part.size
+			});
+			continue;
+		}
+		const data = (await opts.gmail.users.messages.attachments.get({
+			userId: "me",
+			messageId: opts.messageId,
+			id: part.attachmentId
+		})).data.data;
+		if (!data) continue;
+		const buffer = Buffer.from(data, "base64url");
+		const storedName = `${opts.messageId}__${sanitizeFilename(part.filename)}`;
+		const markdownName = `${storedName}.md`;
+		fs.writeFileSync(path.join(attachmentsDir, storedName), buffer);
+		const { markdown } = await convertAttachment(buffer, part.filename, part.mimeType);
+		const mdBody = `# ${part.filename}\n\n_Source: ${opts.source} · ${opts.date}_\n\n${markdown}\n`;
+		fs.writeFileSync(path.join(attachmentsDir, markdownName), mdBody);
+		const ref = `${opts.source}#att:${part.filename}`;
+		const chunks = chunkText(markdown);
+		for (let i = 0; i < chunks.length; i++) await indexInLanceDB(opts.dataDir, opts.slug, chunks[i], `${ref}#${i}`, {
+			date: opts.date,
+			type: "attachment"
+		}).catch((err) => {
+			logger.error("gmail-sync", "attachment index failed", { error: err.message });
+		});
+		saved.push({
+			originalName: part.filename,
+			storedName,
+			markdownName,
+			ref,
+			chunks: chunks.length
+		});
+	} catch (err) {
+		logger.warn("gmail-sync", "attachment failed", {
+			filename: part.filename,
+			error: err.message
+		});
+	}
+	return saved;
+}
+//#endregion
+export { processMessageAttachments };
+
+//# sourceMappingURL=attachments-rLa96rOK.js.map

package/dist/attachments-rLa96rOK.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"attachments-rLa96rOK.js","names":[],"sources":["../src/sync/converters/text.ts","../src/sync/converters/docx.ts","../src/sync/converters/xlsx.ts","../src/sync/converters/pptx.ts","../src/sync/converters/pdf.ts","../src/sync/converters/image.ts","../src/sync/converters/registry.ts","../src/sync/attachments.ts"],"sourcesContent":["// src/sync/converters/text.ts\nimport type { Converter, ConversionResult } from \"./types.js\";\n\n/** Escape a CSV cell for safe inclusion in a Markdown table cell. */\nfunction mdCell(value: string): string {\n  return value.replace(/\\\\/g, \"\\\\\\\\\").replace(/\\|/g, \"\\\\|\").replace(/\\r?\\n/g, \" \").trim();\n}\n\n/**\n * Minimal RFC-4180-ish CSV line splitter: handles quoted fields containing\n * commas and escaped double quotes. Good enough for rendering CSV attachments\n * as readable Markdown tables (we are not round-tripping data).\n */\nexport function parseCsvLine(line: string): string[] {\n  const out: string[] = [];\n  let cur = \"\";\n  let inQuotes = false;\n  for (let i = 0; i < line.length; i++) {\n    const ch = line[i];\n    if (inQuotes) {\n      if (ch === '\"') {\n        if (line[i + 1] === '\"') {\n          cur += '\"';\n          i++;\n        } else {\n          inQuotes = false;\n        }\n      } else {\n        cur += ch;\n      }\n    } else if (ch === '\"') {\n      inQuotes = true;\n    } else if (ch === \",\") {\n      out.push(cur);\n      cur = \"\";\n    } else {\n      cur += ch;\n    }\n  }\n  out.push(cur);\n  return out;\n}\n\n/** Render a matrix of cells as a GitHub-flavored Markdown pipe table. */\nexport function rowsToMarkdown(rows: string[][]): string {\n  if (rows.length === 0) return \"\";\n  const width = Math.max(...rows.map((r) => r.length));\n  const pad = (r: string[]): string[] =>\n    Array.from({ length: width }, (_, i) => mdCell(r[i] ?? \"\"));\n\n  const header = pad(rows[0] ?? []);\n  const lines = [\n    `| ${header.join(\" | \")} |`,\n    `| ${header.map(() => \"---\").join(\" | \")} |`,\n    ...rows.slice(1).map((r) => `| ${pad(r).join(\" | \")} |`),\n  ];\n  return lines.join(\"\\n\");\n}\n\n/** Render CSV text as a GitHub-flavored Markdown pipe table. */\nexport function csvToMarkdown(csv: string): string {\n  const rows = csv\n    .split(/\\r?\\n/)\n    .filter((l) => l.trim().length > 0)\n    .map(parseCsvLine);\n  return rowsToMarkdown(rows);\n}\n\nconst TEXT_EXTENSIONS = [\"txt\", \"text\", \"log\", \"md\", \"markdown\"];\nconst CODE_FENCE_EXTENSIONS: Record<string, string> = {\n  json: \"json\",\n  xml: \"xml\",\n  yaml: \"yaml\",\n  yml: \"yaml\",\n};\n\n/**\n * Converter for plain-text-ish attachments: Markdown/text passthrough, CSV/TSV\n * to Markdown tables, and structured text (JSON/XML/YAML) into fenced code\n * blocks so they stay readable and searchable without a heavy parser.\n */\nexport const textConverter: Converter = {\n  name: \"text\",\n  extensions: [...TEXT_EXTENSIONS, \"csv\", \"tsv\", ...Object.keys(CODE_FENCE_EXTENSIONS)],\n  mimeTypes: [\"text/plain\", \"text/csv\", \"text/markdown\", \"application/json\", \"text/*\"],\n  convert(buffer: Buffer, filename: string): Promise<ConversionResult> {\n    const ext = filename.split(\".\").pop()?.toLowerCase() ?? \"\";\n    const content = buffer.toString(\"utf-8\");\n\n    if (ext === \"csv\") {\n      return Promise.resolve({ markdown: csvToMarkdown(content), meta: { format: \"csv\" } });\n    }\n    if (ext === \"tsv\") {\n      const asCsv = content.replace(/\\t/g, \",\");\n      return Promise.resolve({ markdown: csvToMarkdown(asCsv), meta: { format: \"tsv\" } });\n    }\n    const fence = CODE_FENCE_EXTENSIONS[ext];\n    if (fence) {\n      return Promise.resolve({\n        markdown: `\\`\\`\\`${fence}\\n${content.trim()}\\n\\`\\`\\``,\n        meta: { format: fence },\n      });\n    }\n    // Markdown / plain text: pass through verbatim.\n    return Promise.resolve({ markdown: content.trim(), meta: { format: \"text\" } });\n  },\n};\n","// src/sync/converters/docx.ts\nimport type { Converter, ConversionResult } from \"./types.js\";\nimport { htmlToMarkdown } from \"./html.js\";\n\n/**\n * DOCX → Markdown via mammoth (DOCX → semantic HTML) then Turndown (HTML →\n * Markdown). Mammoth's own Markdown output is deprecated; the HTML route keeps\n * tables, lists and headings intact. mammoth is loaded lazily.\n */\nexport const docxConverter: Converter = {\n  name: \"docx\",\n  extensions: [\"docx\"],\n  mimeTypes: [\"application/vnd.openxmlformats-officedocument.wordprocessingml.document\"],\n  async convert(buffer: Buffer): Promise<ConversionResult> {\n    const mammoth = (await import(\"mammoth\")).default;\n    const { value: html, messages } = await mammoth.convertToHtml({ buffer });\n    const markdown = await htmlToMarkdown(html);\n    return {\n      markdown,\n      meta: { format: \"docx\", warnings: messages.filter((m) => m.type === \"warning\").length },\n    };\n  },\n};\n","// src/sync/converters/xlsx.ts\nimport type { Converter, ConversionResult } from \"./types.js\";\nimport { rowsToMarkdown } from \"./text.js\";\n\n/** Render a single ExcelJS cell value as plain text. */\nfunction cellText(value: unknown): string {\n  if (value === null || value === undefined) return \"\";\n  if (value instanceof Date) return value.toISOString().slice(0, 10);\n  if (typeof value === \"object\") {\n    const v = value as Record<string, unknown>;\n    if (typeof v[\"text\"] === \"string\") return v[\"text\"];\n    if (\"result\" in v) return String(v[\"result\"] ?? \"\");\n    if (Array.isArray(v[\"richText\"])) {\n      return (v[\"richText\"] as Array<{ text?: string }>).map((r) => r.text ?? \"\").join(\"\");\n    }\n    if (\"hyperlink\" in v) return String(v[\"text\"] ?? v[\"hyperlink\"] ?? \"\");\n  }\n  return String(value);\n}\n\n/**\n * Spreadsheet (XLSX) → Markdown via ExcelJS. Each worksheet becomes a\n * `## <sheet name>` section followed by a GitHub-flavored Markdown table.\n * ExcelJS is loaded lazily.\n */\nexport const xlsxConverter: Converter = {\n  name: \"xlsx\",\n  extensions: [\"xlsx\", \"xlsm\"],\n  mimeTypes: [\n    \"application/vnd.openxmlformats-officedocument.spreadsheetml.sheet\",\n    \"application/vnd.ms-excel\",\n  ],\n  async convert(buffer: Buffer): Promise<ConversionResult> {\n    const ExcelJS = (await import(\"exceljs\")).default;\n    const wb = new ExcelJS.Workbook();\n    // ExcelJS's typings predate the @types/node generic Buffer; widen via ArrayBuffer.\n    await wb.xlsx.load(buffer as unknown as ArrayBuffer);\n\n    const sections: string[] = [];\n    const sheetNames: string[] = [];\n    wb.eachSheet((sheet) => {\n      sheetNames.push(sheet.name);\n      const rows: string[][] = [];\n      sheet.eachRow({ includeEmpty: false }, (row) => {\n        const cells: string[] = [];\n        row.eachCell({ includeEmpty: true }, (cell) => {\n          cells.push(cellText(cell.value));\n        });\n        rows.push(cells);\n      });\n      const table = rowsToMarkdown(rows);\n      if (table) sections.push(`## ${sheet.name}\\n\\n${table}`);\n    });\n\n    return { markdown: sections.join(\"\\n\\n\"), meta: { format: \"xlsx\", sheets: sheetNames } };\n  },\n};\n","// src/sync/converters/pptx.ts\nimport type { Converter, ConversionResult } from \"./types.js\";\n\n/** Extract the visible text runs (`<a:t>…</a:t>`) from one slide's XML. */\nexport function extractSlideText(xml: string): string {\n  const runs = [...xml.matchAll(/<a:t>([\\s\\S]*?)<\\/a:t>/g)].map((m) =>\n    (m[1] ?? \"\")\n      .replace(/&amp;/g, \"&\")\n      .replace(/&lt;/g, \"<\")\n      .replace(/&gt;/g, \">\")\n      .replace(/&quot;/g, '\"')\n      .replace(/&apos;/g, \"'\")\n  );\n  return runs.join(\" \").replace(/\\s+/g, \" \").trim();\n}\n\nfunction slideNumber(entryName: string): number {\n  const m = entryName.match(/slide(\\d+)\\.xml$/);\n  return m ? parseInt(m[1] ?? \"0\", 10) : 0;\n}\n\n/**\n * PPTX → Markdown. A .pptx is a zip; slide text lives in `ppt/slides/slideN.xml`\n * as `<a:t>` runs. We unzip with adm-zip (already a dependency) and emit one\n * `## Slide N` section per slide — no extra native parser needed.\n */\nexport const pptxConverter: Converter = {\n  name: \"pptx\",\n  extensions: [\"pptx\"],\n  mimeTypes: [\"application/vnd.openxmlformats-officedocument.presentationml.presentation\"],\n  async convert(buffer: Buffer): Promise<ConversionResult> {\n    const AdmZip = (await import(\"adm-zip\")).default;\n    const zip = new AdmZip(buffer);\n    const slides = zip\n      .getEntries()\n      .filter((e) => /^ppt\\/slides\\/slide\\d+\\.xml$/.test(e.entryName))\n      .sort((a, b) => slideNumber(a.entryName) - slideNumber(b.entryName));\n\n    const sections: string[] = [];\n    for (const entry of slides) {\n      const text = extractSlideText(entry.getData().toString(\"utf-8\"));\n      if (text) sections.push(`## Slide ${slideNumber(entry.entryName)}\\n\\n${text}`);\n    }\n    return { markdown: sections.join(\"\\n\\n\"), meta: { format: \"pptx\", slides: slides.length } };\n  },\n};\n","// src/sync/converters/pdf.ts\nimport type { Converter, ConversionResult } from \"./types.js\";\n\n/** Whether scanned-PDF OCR is enabled (opt-in, requires @napi-rs/canvas). */\nexport function isPdfOcrEnabled(): boolean {\n  const v = (process.env[\"DXCRM_PDF_OCR\"] ?? \"\").toLowerCase();\n  return v === \"1\" || v === \"true\" || v === \"yes\" || v === \"on\";\n}\n\n/**\n * OCR a scanned PDF by rendering each page to a PNG (unpdf + @napi-rs/canvas)\n * and running tesseract.js over it. The canvas backend is an optional peer\n * dependency; if it isn't installed we throw a clear, catchable error so the\n * caller can fall back to the OCR-candidate stub. Page count is capped via\n * DXCRM_PDF_OCR_MAX_PAGES (default 20) to keep this slow path bounded.\n */\nasync function ocrPdf(pdf: unknown, totalPages: number): Promise<string> {\n  // Fail fast with a helpful message when the optional canvas backend is absent.\n  try {\n    await import(\"@napi-rs/canvas\");\n  } catch {\n    throw new Error(\n      \"PDF OCR requires the optional '@napi-rs/canvas' package — run `npm install @napi-rs/canvas`\"\n    );\n  }\n\n  const { renderPageAsImage } = await import(\"unpdf\");\n  const { recognize } = await import(\"tesseract.js\");\n  const lang = process.env[\"DXCRM_OCR_LANG\"] ?? \"eng\";\n  const canvasImport = (): Promise<unknown> => import(\"@napi-rs/canvas\");\n\n  const maxPages = Number(process.env[\"DXCRM_PDF_OCR_MAX_PAGES\"] ?? 20) || 20;\n  const pageCount = Math.min(totalPages, maxPages);\n\n  const pages: string[] = [];\n  for (let p = 1; p <= pageCount; p++) {\n    const png = await renderPageAsImage(pdf as never, p, { canvasImport, scale: 2 } as never);\n    const {\n      data: { text },\n    } = await recognize(Buffer.from(png), lang);\n    if (text.trim()) pages.push(`## Page ${p}\\n\\n${text.trim()}`);\n  }\n  return pages.join(\"\\n\\n\");\n}\n\n/**\n * PDF → Markdown. Extracts the digital text layer with unpdf (a serverless\n * pdf.js build). Scanned PDFs have no text layer: when DXCRM_PDF_OCR is enabled\n * they are rendered and OCR'd page-by-page, otherwise they're flagged as OCR\n * candidates rather than emitting garbage. unpdf is loaded lazily.\n */\nexport const pdfConverter: Converter = {\n  name: \"pdf\",\n  extensions: [\"pdf\"],\n  mimeTypes: [\"application/pdf\"],\n  async convert(buffer: Buffer, filename: string): Promise<ConversionResult> {\n    const { extractText, getDocumentProxy } = await import(\"unpdf\");\n    const pdf = await getDocumentProxy(new Uint8Array(buffer));\n    const { totalPages, text } = await extractText(pdf, { mergePages: true });\n    const merged = (Array.isArray(text) ? text.join(\"\\n\\n\") : text).trim();\n\n    if (merged) return { markdown: merged, meta: { format: \"pdf\", pages: totalPages } };\n\n    // No text layer — scanned PDF. OCR if opted in, else flag as a candidate.\n    if (isPdfOcrEnabled()) {\n      try {\n        const ocr = await ocrPdf(pdf, totalPages);\n        if (ocr.trim()) {\n          return {\n            markdown: `> _OCR of \\`${filename}\\` (${totalPages} pages):_\\n\\n${ocr.trim()}`,\n            meta: { format: \"pdf\", pages: totalPages, ocr: true },\n          };\n        }\n      } catch (err) {\n        process.stderr.write(\n          `[converters] pdf OCR failed for ${filename}: ${(err as Error).message}\\n`\n        );\n      }\n    }\n\n    return { markdown: \"\", meta: { format: \"pdf\", pages: totalPages, ocrCandidate: true } };\n  },\n};\n","// src/sync/converters/image.ts\nimport type { Converter, ConversionResult } from \"./types.js\";\n\n/**\n * Image → Markdown via Tesseract.js OCR (pure-JS, 100+ languages, fully local).\n * This is the heaviest converter: tesseract.js downloads a WASM core and\n * language data on first use, so it is loaded lazily and only invoked for image\n * attachments. Language defaults to English, override with DXCRM_OCR_LANG.\n */\nexport const imageConverter: Converter = {\n  name: \"image\",\n  extensions: [\"png\", \"jpg\", \"jpeg\", \"tif\", \"tiff\", \"bmp\", \"webp\", \"gif\", \"pbm\"],\n  mimeTypes: [\"image/*\"],\n  async convert(buffer: Buffer, filename: string): Promise<ConversionResult> {\n    const lang = process.env[\"DXCRM_OCR_LANG\"] ?? \"eng\";\n    const { recognize } = await import(\"tesseract.js\");\n    const {\n      data: { text },\n    } = await recognize(buffer, lang);\n    const ocr = text.trim();\n    return {\n      markdown: ocr ? `> _OCR of \\`${filename}\\`:_\\n\\n${ocr}` : \"\",\n      meta: { format: \"image\", ocr: true, lang },\n    };\n  },\n};\n","// src/sync/converters/registry.ts\nimport type { Converter, ConversionResult } from \"./types.js\";\nimport { textConverter } from \"./text.js\";\nimport { docxConverter } from \"./docx.js\";\nimport { xlsxConverter } from \"./xlsx.js\";\nimport { pptxConverter } from \"./pptx.js\";\nimport { pdfConverter } from \"./pdf.js\";\nimport { htmlConverter } from \"./html.js\";\nimport { imageConverter } from \"./image.js\";\n\n/**\n * Fallback converter for unknown/binary attachments: emit a small metadata stub\n * instead of garbage bytes, so the attachment is still recorded and linkable.\n */\nexport const fallbackConverter: Converter = {\n  name: \"binary\",\n  extensions: [],\n  convert(buffer: Buffer, filename: string): Promise<ConversionResult> {\n    const kb = Math.max(1, Math.round(buffer.length / 1024));\n    return Promise.resolve({\n      markdown: `> _Binary attachment \\`${filename}\\` (${kb} KB) — no text representation available._`,\n      meta: { format: \"binary\", bytes: buffer.length },\n    });\n  },\n};\n\n/**\n * Ordered converter registry. Earlier entries win on extension conflicts. The\n * text converter is intentionally last among the \"real\" converters so that more\n * specific formats (html, etc.) take precedence over generic text matching.\n */\nexport const CONVERTERS: Converter[] = [\n  docxConverter,\n  xlsxConverter,\n  pptxConverter,\n  pdfConverter,\n  imageConverter,\n  htmlConverter,\n  textConverter,\n];\n\nfunction extensionOf(filename: string): string {\n  return filename.split(\".\").pop()?.toLowerCase() ?? \"\";\n}\n\nfunction mimeMatches(converter: Converter, mime: string): boolean {\n  if (!converter.mimeTypes) return false;\n  const lower = mime.toLowerCase();\n  return converter.mimeTypes.some((m) => {\n    const ml = m.toLowerCase();\n    if (ml.endsWith(\"/*\")) return lower.startsWith(ml.slice(0, -1));\n    return ml === lower;\n  });\n}\n\n/**\n * Pick the converter for an attachment by file extension first (most reliable\n * for Gmail attachments, which always carry a filename), then by MIME type.\n * Returns `undefined` when nothing matches.\n */\nexport function matchConverter(filename: string, mime?: string): Converter | undefined {\n  const ext = extensionOf(filename);\n  if (ext) {\n    const byExt = CONVERTERS.find((c) => c.extensions.includes(ext));\n    if (byExt) return byExt;\n  }\n  if (mime) {\n    const byMime = CONVERTERS.find((c) => mimeMatches(c, mime));\n    if (byMime) return byMime;\n  }\n  return undefined;\n}\n\n/**\n * Convert an attachment to Markdown, dispatching to the best converter and\n * falling back to a metadata stub. Converter errors never throw: they are\n * swallowed into the fallback so a single bad attachment can't break a sync.\n */\nexport async function convertAttachment(\n  buffer: Buffer,\n  filename: string,\n  mime?: string\n): Promise<ConversionResult> {\n  const converter = matchConverter(filename, mime) ?? fallbackConverter;\n  try {\n    const result = await converter.convert(buffer, filename);\n    if (!result.markdown.trim()) {\n      return {\n        markdown: `> _Attachment \\`${filename}\\` contained no extractable text._`,\n        meta: { ...result.meta, empty: true },\n      };\n    }\n    return result;\n  } catch (err) {\n    process.stderr.write(\n      `[converters] ${converter.name} failed for ${filename}: ${(err as Error).message}\\n`\n    );\n    return fallbackConverter.convert(buffer, filename);\n  }\n}\n","// src/sync/attachments.ts\nimport fs from \"fs\";\nimport path from \"path\";\nimport type { gmail_v1 } from \"@googleapis/gmail\";\nimport { convertAttachment } from \"./converters/registry.js\";\nimport { chunkText } from \"../core/chunk.js\";\nimport { assertSafeSlug } from \"../fs/customer-dir.js\";\nimport { logger } from \"../core/logger.js\";\n\n/** Default per-attachment size cap (skip larger blobs to keep syncs bounded). */\nexport const DEFAULT_MAX_ATTACHMENT_BYTES = 25 * 1024 * 1024;\n\nexport interface AttachmentPart {\n  filename: string;\n  mimeType: string;\n  attachmentId: string;\n  size: number;\n}\n\nexport interface SavedAttachment {\n  /** Original filename as sent. */\n  originalName: string;\n  /** Stored raw filename (sanitized, message-prefixed) under attachments/. */\n  storedName: string;\n  /** Markdown filename under attachments/. */\n  markdownName: string;\n  /** Source ref used for LanceDB indexing. */\n  ref: string;\n  /** Number of indexed chunks produced from the Markdown. */\n  chunks: number;\n}\n\n/**\n * Recursively collect downloadable attachment parts from a Gmail message\n * payload — any MIME part that carries both a filename and a body.attachmentId.\n * Inline parts without a filename (e.g. signature logos) are ignored.\n */\nexport function collectAttachmentParts(\n  payload: gmail_v1.Schema$MessagePart | undefined\n): AttachmentPart[] {\n  const out: AttachmentPart[] = [];\n  const walk = (part?: gmail_v1.Schema$MessagePart): void => {\n    if (!part) return;\n    const filename = part.filename ?? \"\";\n    const attachmentId = part.body?.attachmentId ?? \"\";\n    if (filename && attachmentId) {\n      out.push({\n        filename,\n        mimeType: part.mimeType ?? \"application/octet-stream\",\n        attachmentId,\n        size: part.body?.size ?? 0,\n      });\n    }\n    for (const child of part.parts ?? []) walk(child);\n  };\n  walk(payload);\n  return out;\n}\n\n/** Make a filename safe for use as a single path segment. */\nexport function sanitizeFilename(name: string): string {\n  const base = name.split(/[\\\\/]/).pop() ?? name;\n  return (\n    base\n      .replace(/[^a-zA-Z0-9._-]+/g, \"_\")\n      .replace(/^_+|_+$/g, \"\")\n      .slice(0, 120) || \"attachment\"\n  );\n}\n\n/**\n * Download, convert and index every attachment of a single Gmail message.\n *\n * For each attachment: the raw bytes are saved under\n * `customers/<slug>/attachments/<messageId>__<name>`, converted to a sibling\n * `.md` file, and the Markdown is chunked and indexed into LanceDB so the\n * attachment's content is semantically searchable. Failures on a single\n * attachment are logged and skipped — they never abort the message sync.\n */\nexport async function processMessageAttachments(opts: {\n  gmail: gmail_v1.Gmail;\n  dataDir: string;\n  slug: string;\n  messageId: string;\n  source: string;\n  payload: gmail_v1.Schema$MessagePart | undefined;\n  date: string;\n  maxBytes?: number;\n}): Promise<SavedAttachment[]> {\n  const parts = collectAttachmentParts(opts.payload);\n  if (parts.length === 0) return [];\n\n  assertSafeSlug(opts.slug);\n  const maxBytes = opts.maxBytes ?? DEFAULT_MAX_ATTACHMENT_BYTES;\n  const attachmentsDir = path.join(opts.dataDir, \"customers\", opts.slug, \"attachments\");\n  fs.mkdirSync(attachmentsDir, { recursive: true });\n\n  const { indexInLanceDB } = await import(\"../core/lancedb.js\");\n  const saved: SavedAttachment[] = [];\n\n  for (const part of parts) {\n    try {\n      if (part.size > maxBytes) {\n        logger.warn(\"gmail-sync\", \"skipping oversized attachment\", {\n          filename: part.filename,\n          bytes: part.size,\n        });\n        continue;\n      }\n\n      const resp = await opts.gmail.users.messages.attachments.get({\n        userId: \"me\",\n        messageId: opts.messageId,\n        id: part.attachmentId,\n      });\n      const data = resp.data.data;\n      if (!data) continue;\n      const buffer = Buffer.from(data, \"base64url\");\n\n      const storedName = `${opts.messageId}__${sanitizeFilename(part.filename)}`;\n      const markdownName = `${storedName}.md`;\n      fs.writeFileSync(path.join(attachmentsDir, storedName), buffer);\n\n      const { markdown } = await convertAttachment(buffer, part.filename, part.mimeType);\n      const mdBody = `# ${part.filename}\\n\\n_Source: ${opts.source} · ${opts.date}_\\n\\n${markdown}\\n`;\n      fs.writeFileSync(path.join(attachmentsDir, markdownName), mdBody);\n\n      const ref = `${opts.source}#att:${part.filename}`;\n      const chunks = chunkText(markdown);\n      for (let i = 0; i < chunks.length; i++) {\n        await indexInLanceDB(opts.dataDir, opts.slug, chunks[i]!, `${ref}#${i}`, {\n          date: opts.date,\n          type: \"attachment\",\n        }).catch((err: unknown) => {\n          logger.error(\"gmail-sync\", \"attachment index failed\", {\n            error: (err as Error).message,\n          });\n        });\n      }\n\n      saved.push({\n        originalName: part.filename,\n        storedName,\n        markdownName,\n        ref,\n        chunks: chunks.length,\n      });\n    } catch (err) {\n      logger.warn(\"gmail-sync\", \"attachment failed\", {\n        filename: part.filename,\n        error: (err as Error).message,\n      });\n    }\n  }\n\n  return saved;\n}\n"],"mappings":";;;;;;;;AAIA,SAAS,OAAO,OAAuB;CACrC,OAAO,MAAM,QAAQ,OAAO,MAAM,EAAE,QAAQ,OAAO,KAAK,EAAE,QAAQ,UAAU,GAAG,EAAE,KAAK;AACxF;;;;;;AAOA,SAAgB,aAAa,MAAwB;CACnD,MAAM,MAAgB,CAAC;CACvB,IAAI,MAAM;CACV,IAAI,WAAW;CACf,KAAK,IAAI,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;EACpC,MAAM,KAAK,KAAK;EAChB,IAAI,UACF,IAAI,OAAO,MACT,IAAI,KAAK,IAAI,OAAO,MAAK;GACvB,OAAO;GACP;EACF,OACE,WAAW;OAGb,OAAO;OAEJ,IAAI,OAAO,MAChB,WAAW;OACN,IAAI,OAAO,KAAK;GACrB,IAAI,KAAK,GAAG;GACZ,MAAM;EACR,OACE,OAAO;CAEX;CACA,IAAI,KAAK,GAAG;CACZ,OAAO;AACT;;AAGA,SAAgB,eAAe,MAA0B;CACvD,IAAI,KAAK,WAAW,GAAG,OAAO;CAC9B,MAAM,QAAQ,KAAK,IAAI,GAAG,KAAK,KAAK,MAAM,EAAE,MAAM,CAAC;CACnD,MAAM,OAAO,MACX,MAAM,KAAK,EAAE,QAAQ,MAAM,IAAI,GAAG,MAAM,OAAO,EAAE,MAAM,EAAE,CAAC;CAE5D,MAAM,SAAS,IAAI,KAAK,MAAM,CAAC,CAAC;CAMhC,OAAO;EAJL,KAAK,OAAO,KAAK,KAAK,EAAE;EACxB,KAAK,OAAO,UAAU,KAAK,EAAE,KAAK,KAAK,EAAE;EACzC,GAAG,KAAK,MAAM,CAAC,EAAE,KAAK,MAAM,KAAK,IAAI,CAAC,EAAE,KAAK,KAAK,EAAE,GAAG;CAE9C,EAAE,KAAK,IAAI;AACxB;;AAGA,SAAgB,cAAc,KAAqB;CAKjD,OAAO,eAJM,IACV,MAAM,OAAO,EACb,QAAQ,MAAM,EAAE,KAAK,EAAE,SAAS,CAAC,EACjC,IAAI,YACkB,CAAC;AAC5B;AAEA,MAAM,kBAAkB;CAAC;CAAO;CAAQ;CAAO;CAAM;AAAU;AAC/D,MAAM,wBAAgD;CACpD,MAAM;CACN,KAAK;CACL,MAAM;CACN,KAAK;AACP;;;;;;AAOA,MAAa,gBAA2B;CACtC,MAAM;CACN,YAAY;EAAC,GAAG;EAAiB;EAAO;EAAO,GAAG,OAAO,KAAK,qBAAqB;CAAC;CACpF,WAAW;EAAC;EAAc;EAAY;EAAiB;EAAoB;CAAQ;CACnF,QAAQ,QAAgB,UAA6C;EACnE,MAAM,MAAM,SAAS,MAAM,GAAG,EAAE,IAAI,GAAG,YAAY,KAAK;EACxD,MAAM,UAAU,OAAO,SAAS,OAAO;EAEvC,IAAI,QAAQ,OACV,OAAO,QAAQ,QAAQ;GAAE,UAAU,cAAc,OAAO;GAAG,MAAM,EAAE,QAAQ,MAAM;EAAE,CAAC;EAEtF,IAAI,QAAQ,OAAO;GACjB,MAAM,QAAQ,QAAQ,QAAQ,OAAO,GAAG;GACxC,OAAO,QAAQ,QAAQ;IAAE,UAAU,cAAc,KAAK;IAAG,MAAM,EAAE,QAAQ,MAAM;GAAE,CAAC;EACpF;EACA,MAAM,QAAQ,sBAAsB;EACpC,IAAI,OACF,OAAO,QAAQ,QAAQ;GACrB,UAAU,SAAS,MAAM,IAAI,QAAQ,KAAK,EAAE;GAC5C,MAAM,EAAE,QAAQ,MAAM;EACxB,CAAC;EAGH,OAAO,QAAQ,QAAQ;GAAE,UAAU,QAAQ,KAAK;GAAG,MAAM,EAAE,QAAQ,OAAO;EAAE,CAAC;CAC/E;AACF;;;;;;;;ACjGA,MAAa,gBAA2B;CACtC,MAAM;CACN,YAAY,CAAC,MAAM;CACnB,WAAW,CAAC,yEAAyE;CACrF,MAAM,QAAQ,QAA2C;EAEvD,MAAM,EAAE,OAAO,MAAM,aAAa,OADjB,MAAM,OAAO,YAAY,QACM,cAAc,EAAE,OAAO,CAAC;EAExE,OAAO;GACL,UAAA,MAFqB,eAAe,IAAI;GAGxC,MAAM;IAAE,QAAQ;IAAQ,UAAU,SAAS,QAAQ,MAAM,EAAE,SAAS,SAAS,EAAE;GAAO;EACxF;CACF;AACF;;;;ACjBA,SAAS,SAAS,OAAwB;CACxC,IAAI,UAAU,QAAQ,UAAU,KAAA,GAAW,OAAO;CAClD,IAAI,iBAAiB,MAAM,OAAO,MAAM,YAAY,EAAE,MAAM,GAAG,EAAE;CACjE,IAAI,OAAO,UAAU,UAAU;EAC7B,MAAM,IAAI;EACV,IAAI,OAAO,EAAE,YAAY,UAAU,OAAO,EAAE;EAC5C,IAAI,YAAY,GAAG,OAAO,OAAO,EAAE,aAAa,EAAE;EAClD,IAAI,MAAM,QAAQ,EAAE,WAAW,GAC7B,OAAQ,EAAE,YAAyC,KAAK,MAAM,EAAE,QAAQ,EAAE,EAAE,KAAK,EAAE;EAErF,IAAI,eAAe,GAAG,OAAO,OAAO,EAAE,WAAW,EAAE,gBAAgB,EAAE;CACvE;CACA,OAAO,OAAO,KAAK;AACrB;;;;;;AAOA,MAAa,gBAA2B;CACtC,MAAM;CACN,YAAY,CAAC,QAAQ,MAAM;CAC3B,WAAW,CACT,qEACA,0BACF;CACA,MAAM,QAAQ,QAA2C;EAEvD,MAAM,KAAK,KADM,OAAM,OAAO,aAAY,QACnB,SAAS;EAEhC,MAAM,GAAG,KAAK,KAAK,MAAgC;EAEnD,MAAM,WAAqB,CAAC;EAC5B,MAAM,aAAuB,CAAC;EAC9B,GAAG,WAAW,UAAU;GACtB,WAAW,KAAK,MAAM,IAAI;GAC1B,MAAM,OAAmB,CAAC;GAC1B,MAAM,QAAQ,EAAE,cAAc,MAAM,IAAI,QAAQ;IAC9C,MAAM,QAAkB,CAAC;IACzB,IAAI,SAAS,EAAE,cAAc,KAAK,IAAI,SAAS;KAC7C,MAAM,KAAK,SAAS,KAAK,KAAK,CAAC;IACjC,CAAC;IACD,KAAK,KAAK,KAAK;GACjB,CAAC;GACD,MAAM,QAAQ,eAAe,IAAI;GACjC,IAAI,OAAO,SAAS,KAAK,MAAM,MAAM,KAAK,MAAM,OAAO;EACzD,CAAC;EAED,OAAO;GAAE,UAAU,SAAS,KAAK,MAAM;GAAG,MAAM;IAAE,QAAQ;IAAQ,QAAQ;GAAW;EAAE;CACzF;AACF;;;;ACpDA,SAAgB,iBAAiB,KAAqB;CASpD,OARa,CAAC,GAAG,IAAI,SAAS,yBAAyB,CAAC,EAAE,KAAK,OAC5D,EAAE,MAAM,IACN,QAAQ,UAAU,GAAG,EACrB,QAAQ,SAAS,GAAG,EACpB,QAAQ,SAAS,GAAG,EACpB,QAAQ,WAAW,IAAG,EACtB,QAAQ,WAAW,GAAG,CAEjB,EAAE,KAAK,GAAG,EAAE,QAAQ,QAAQ,GAAG,EAAE,KAAK;AAClD;AAEA,SAAS,YAAY,WAA2B;CAC9C,MAAM,IAAI,UAAU,MAAM,kBAAkB;CAC5C,OAAO,IAAI,SAAS,EAAE,MAAM,KAAK,EAAE,IAAI;AACzC;;;;;;AAOA,MAAa,gBAA2B;CACtC,MAAM;CACN,YAAY,CAAC,MAAM;CACnB,WAAW,CAAC,2EAA2E;CACvF,MAAM,QAAQ,QAA2C;EACvD,MAAM,UAAU,MAAM,OAAO,YAAY;EAEzC,MAAM,SAAS,IADC,OAAO,MACN,EACd,WAAW,EACX,QAAQ,MAAM,+BAA+B,KAAK,EAAE,SAAS,CAAC,EAC9D,MAAM,GAAG,MAAM,YAAY,EAAE,SAAS,IAAI,YAAY,EAAE,SAAS,CAAC;EAErE,MAAM,WAAqB,CAAC;EAC5B,KAAK,MAAM,SAAS,QAAQ;GAC1B,MAAM,OAAO,iBAAiB,MAAM,QAAQ,EAAE,SAAS,OAAO,CAAC;GAC/D,IAAI,MAAM,SAAS,KAAK,YAAY,YAAY,MAAM,SAAS,EAAE,MAAM,MAAM;EAC/E;EACA,OAAO;GAAE,UAAU,SAAS,KAAK,MAAM;GAAG,MAAM;IAAE,QAAQ;IAAQ,QAAQ,OAAO;GAAO;EAAE;CAC5F;AACF;;;;ACzCA,SAAgB,kBAA2B;CACzC,MAAM,KAAK,QAAQ,IAAI,oBAAoB,IAAI,YAAY;CAC3D,OAAO,MAAM,OAAO,MAAM,UAAU,MAAM,SAAS,MAAM;AAC3D;;;;;;;;AASA,eAAe,OAAO,KAAc,YAAqC;CAEvE,IAAI;EACF,MAAM,OAAO;CACf,QAAQ;EACN,MAAM,IAAI,MACR,6FACF;CACF;CAEA,MAAM,EAAE,sBAAsB,MAAM,OAAO;CAC3C,MAAM,EAAE,cAAc,MAAM,OAAO;CACnC,MAAM,OAAO,QAAQ,IAAI,qBAAqB;CAC9C,MAAM,qBAAuC,OAAO;CAEpD,MAAM,WAAW,OAAO,QAAQ,IAAI,8BAA8B,EAAE,KAAK;CACzE,MAAM,YAAY,KAAK,IAAI,YAAY,QAAQ;CAE/C,MAAM,QAAkB,CAAC;CACzB,KAAK,IAAI,IAAI,GAAG,KAAK,WAAW,KAAK;EACnC,MAAM,MAAM,MAAM,kBAAkB,KAAc,GAAG;GAAE;GAAc,OAAO;EAAE,CAAU;EACxF,MAAM,EACJ,MAAM,EAAE,WACN,MAAM,UAAU,OAAO,KAAK,GAAG,GAAG,IAAI;EAC1C,IAAI,KAAK,KAAK,GAAG,MAAM,KAAK,WAAW,EAAE,MAAM,KAAK,KAAK,GAAG;CAC9D;CACA,OAAO,MAAM,KAAK,MAAM;AAC1B;;;;;;;AAQA,MAAa,eAA0B;CACrC,MAAM;CACN,YAAY,CAAC,KAAK;CAClB,WAAW,CAAC,iBAAiB;CAC7B,MAAM,QAAQ,QAAgB,UAA6C;EACzE,MAAM,EAAE,aAAa,qBAAqB,MAAM,OAAO;EACvD,MAAM,MAAM,MAAM,iBAAiB,IAAI,WAAW,MAAM,CAAC;EACzD,MAAM,EAAE,YAAY,SAAS,MAAM,YAAY,KAAK,EAAE,YAAY,KAAK,CAAC;EACxE,MAAM,UAAU,MAAM,QAAQ,IAAI,IAAI,KAAK,KAAK,MAAM,IAAI,MAAM,KAAK;EAErE,IAAI,QAAQ,OAAO;GAAE,UAAU;GAAQ,MAAM;IAAE,QAAQ;IAAO,OAAO;GAAW;EAAE;EAGlF,IAAI,gBAAgB,GAClB,IAAI;GACF,MAAM,MAAM,MAAM,OAAO,KAAK,UAAU;GACxC,IAAI,IAAI,KAAK,GACX,OAAO;IACL,UAAU,eAAe,SAAS,MAAM,WAAW,eAAe,IAAI,KAAK;IAC3E,MAAM;KAAE,QAAQ;KAAO,OAAO;KAAY,KAAK;IAAK;GACtD;EAEJ,SAAS,KAAK;GACZ,QAAQ,OAAO,MACb,mCAAmC,SAAS,IAAK,IAAc,QAAQ,GACzE;EACF;EAGF,OAAO;GAAE,UAAU;GAAI,MAAM;IAAE,QAAQ;IAAO,OAAO;IAAY,cAAc;GAAK;EAAE;CACxF;AACF;;;;;;;;;ACzEA,MAAa,iBAA4B;CACvC,MAAM;CACN,YAAY;EAAC;EAAO;EAAO;EAAQ;EAAO;EAAQ;EAAO;EAAQ;EAAO;CAAK;CAC7E,WAAW,CAAC,SAAS;CACrB,MAAM,QAAQ,QAAgB,UAA6C;EACzE,MAAM,OAAO,QAAQ,IAAI,qBAAqB;EAC9C,MAAM,EAAE,cAAc,MAAM,OAAO;EACnC,MAAM,EACJ,MAAM,EAAE,WACN,MAAM,UAAU,QAAQ,IAAI;EAChC,MAAM,MAAM,KAAK,KAAK;EACtB,OAAO;GACL,UAAU,MAAM,eAAe,SAAS,UAAU,QAAQ;GAC1D,MAAM;IAAE,QAAQ;IAAS,KAAK;IAAM;GAAK;EAC3C;CACF;AACF;;;;;;;ACXA,MAAa,oBAA+B;CAC1C,MAAM;CACN,YAAY,CAAC;CACb,QAAQ,QAAgB,UAA6C;EACnE,MAAM,KAAK,KAAK,IAAI,GAAG,KAAK,MAAM,OAAO,SAAS,IAAI,CAAC;EACvD,OAAO,QAAQ,QAAQ;GACrB,UAAU,0BAA0B,SAAS,MAAM,GAAG;GACtD,MAAM;IAAE,QAAQ;IAAU,OAAO,OAAO;GAAO;EACjD,CAAC;CACH;AACF;;;;;;AAOA,MAAa,aAA0B;CACrC;CACA;CACA;CACA;CACA;CACA;CACA;AACF;AAEA,SAAS,YAAY,UAA0B;CAC7C,OAAO,SAAS,MAAM,GAAG,EAAE,IAAI,GAAG,YAAY,KAAK;AACrD;AAEA,SAAS,YAAY,WAAsB,MAAuB;CAChE,IAAI,CAAC,UAAU,WAAW,OAAO;CACjC,MAAM,QAAQ,KAAK,YAAY;CAC/B,OAAO,UAAU,UAAU,MAAM,MAAM;EACrC,MAAM,KAAK,EAAE,YAAY;EACzB,IAAI,GAAG,SAAS,IAAI,GAAG,OAAO,MAAM,WAAW,GAAG,MAAM,GAAG,EAAE,CAAC;EAC9D,OAAO,OAAO;CAChB,CAAC;AACH;;;;;;AAOA,SAAgB,eAAe,UAAkB,MAAsC;CACrF,MAAM,MAAM,YAAY,QAAQ;CAChC,IAAI,KAAK;EACP,MAAM,QAAQ,WAAW,MAAM,MAAM,EAAE,WAAW,SAAS,GAAG,CAAC;EAC/D,IAAI,OAAO,OAAO;CACpB;CACA,IAAI,MAAM;EACR,MAAM,SAAS,WAAW,MAAM,MAAM,YAAY,GAAG,IAAI,CAAC;EAC1D,IAAI,QAAQ,OAAO;CACrB;AAEF;;;;;;AAOA,eAAsB,kBACpB,QACA,UACA,MAC2B;CAC3B,MAAM,YAAY,eAAe,UAAU,IAAI,KAAK;CACpD,IAAI;EACF,MAAM,SAAS,MAAM,UAAU,QAAQ,QAAQ,QAAQ;EACvD,IAAI,CAAC,OAAO,SAAS,KAAK,GACxB,OAAO;GACL,UAAU,mBAAmB,SAAS;GACtC,MAAM;IAAE,GAAG,OAAO;IAAM,OAAO;GAAK;EACtC;EAEF,OAAO;CACT,SAAS,KAAK;EACZ,QAAQ,OAAO,MACb,gBAAgB,UAAU,KAAK,cAAc,SAAS,IAAK,IAAc,QAAQ,GACnF;EACA,OAAO,kBAAkB,QAAQ,QAAQ,QAAQ;CACnD;AACF;;;;;;AC9DA,SAAgB,uBACd,SACkB;CAClB,MAAM,MAAwB,CAAC;CAC/B,MAAM,QAAQ,SAA6C;EACzD,IAAI,CAAC,MAAM;EACX,MAAM,WAAW,KAAK,YAAY;EAClC,MAAM,eAAe,KAAK,MAAM,gBAAgB;EAChD,IAAI,YAAY,cACd,IAAI,KAAK;GACP;GACA,UAAU,KAAK,YAAY;GAC3B;GACA,MAAM,KAAK,MAAM,QAAQ;EAC3B,CAAC;EAEH,KAAK,MAAM,SAAS,KAAK,SAAS,CAAC,GAAG,KAAK,KAAK;CAClD;CACA,KAAK,OAAO;CACZ,OAAO;AACT;;AAGA,SAAgB,iBAAiB,MAAsB;CAErD,QADa,KAAK,MAAM,OAAO,EAAE,IAAI,KAAK,MAGrC,QAAQ,qBAAqB,GAAG,EAChC,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,GAAG,KAAK;AAExB;;;;;;;;;;AAWA,eAAsB,0BAA0B,MASjB;CAC7B,MAAM,QAAQ,uBAAuB,KAAK,OAAO;CACjD,IAAI,MAAM,WAAW,GAAG,OAAO,CAAC;CAEhC,eAAe,KAAK,IAAI;CACxB,MAAM,WAAW,KAAK,YAAA;CACtB,MAAM,iBAAiB,KAAK,KAAK,KAAK,SAAS,aAAa,KAAK,MAAM,aAAa;CACpF,GAAG,UAAU,gBAAgB,EAAE,WAAW,KAAK,CAAC;CAEhD,MAAM,EAAE,mBAAmB,MAAM,OAAO;CACxC,MAAM,QAA2B,CAAC;CAElC,KAAK,MAAM,QAAQ,OACjB,IAAI;EACF,IAAI,KAAK,OAAO,UAAU;GACxB,OAAO,KAAK,cAAc,iCAAiC;IACzD,UAAU,KAAK;IACf,OAAO,KAAK;GACd,CAAC;GACD;EACF;EAOA,MAAM,QAAO,MALM,KAAK,MAAM,MAAM,SAAS,YAAY,IAAI;GAC3D,QAAQ;GACR,WAAW,KAAK;GAChB,IAAI,KAAK;EACX,CAAC,GACiB,KAAK;EACvB,IAAI,CAAC,MAAM;EACX,MAAM,SAAS,OAAO,KAAK,MAAM,WAAW;EAE5C,MAAM,aAAa,GAAG,KAAK,UAAU,IAAI,iBAAiB,KAAK,QAAQ;EACvE,MAAM,eAAe,GAAG,WAAW;EACnC,GAAG,cAAc,KAAK,KAAK,gBAAgB,UAAU,GAAG,MAAM;EAE9D,MAAM,EAAE,aAAa,MAAM,kBAAkB,QAAQ,KAAK,UAAU,KAAK,QAAQ;EACjF,MAAM,SAAS,KAAK,KAAK,SAAS,eAAe,KAAK,OAAO,KAAK,KAAK,KAAK,OAAO,SAAS;EAC5F,GAAG,cAAc,KAAK,KAAK,gBAAgB,YAAY,GAAG,MAAM;EAEhE,MAAM,MAAM,GAAG,KAAK,OAAO,OAAO,KAAK;EACvC,MAAM,SAAS,UAAU,QAAQ;EACjC,KAAK,IAAI,IAAI,GAAG,IAAI,OAAO,QAAQ,KACjC,MAAM,eAAe,KAAK,SAAS,KAAK,MAAM,OAAO,IAAK,GAAG,IAAI,GAAG,KAAK;GACvE,MAAM,KAAK;GACX,MAAM;EACR,CAAC,EAAE,OAAO,QAAiB;GACzB,OAAO,MAAM,cAAc,2BAA2B,EACpD,OAAQ,IAAc,QACxB,CAAC;EACH,CAAC;EAGH,MAAM,KAAK;GACT,cAAc,KAAK;GACnB;GACA;GACA;GACA,QAAQ,OAAO;EACjB,CAAC;CACH,SAAS,KAAK;EACZ,OAAO,KAAK,cAAc,qBAAqB;GAC7C,UAAU,KAAK;GACf,OAAQ,IAAc;EACxB,CAAC;CACH;CAGF,OAAO;AACT"}

package/dist/chunk-BfDYWZQ8.cjs

@@ -0,0 +1,32 @@
+//#region src/core/chunk.ts
+/**
+* Split long text into overlapping chunks for embedding/indexing. Each chunk is
+* at most `maxChars`; chunks overlap by `overlap` characters so a query that
+* straddles a boundary still matches. Splits prefer the nearest whitespace
+* before the limit to avoid cutting words mid-token. Short text returns a single
+* trimmed chunk; empty/whitespace-only text returns no chunks.
+*/
+function chunkText(text, options = {}) {
+	const maxChars = options.maxChars ?? 1500;
+	const overlap = Math.min(options.overlap ?? 150, Math.floor(maxChars / 2));
+	const trimmed = text.trim();
+	if (!trimmed) return [];
+	if (trimmed.length <= maxChars) return [trimmed];
+	const chunks = [];
+	let start = 0;
+	while (start < trimmed.length) {
+		let end = Math.min(start + maxChars, trimmed.length);
+		if (end < trimmed.length) {
+			const lastSpace = trimmed.lastIndexOf(" ", end);
+			if (lastSpace > start + maxChars / 2) end = lastSpace;
+		}
+		chunks.push(trimmed.slice(start, end).trim());
+		if (end >= trimmed.length) break;
+		start = end - overlap;
+	}
+	return chunks.filter((c) => c.length > 0);
+}
+//#endregion
+exports.chunkText = chunkText;
+
+//# sourceMappingURL=chunk-BfDYWZQ8.cjs.map

package/dist/chunk-BfDYWZQ8.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"chunk-BfDYWZQ8.cjs","names":[],"sources":["../src/core/chunk.ts"],"sourcesContent":["// src/core/chunk.ts\n\nexport interface ChunkOptions {\n  /** Target maximum characters per chunk. */\n  maxChars?: number;\n  /** Characters of overlap carried from the end of one chunk into the next. */\n  overlap?: number;\n}\n\n/**\n * Split long text into overlapping chunks for embedding/indexing. Each chunk is\n * at most `maxChars`; chunks overlap by `overlap` characters so a query that\n * straddles a boundary still matches. Splits prefer the nearest whitespace\n * before the limit to avoid cutting words mid-token. Short text returns a single\n * trimmed chunk; empty/whitespace-only text returns no chunks.\n */\nexport function chunkText(text: string, options: ChunkOptions = {}): string[] {\n  const maxChars = options.maxChars ?? 1500;\n  const overlap = Math.min(options.overlap ?? 150, Math.floor(maxChars / 2));\n  const trimmed = text.trim();\n  if (!trimmed) return [];\n  if (trimmed.length <= maxChars) return [trimmed];\n\n  const chunks: string[] = [];\n  let start = 0;\n  while (start < trimmed.length) {\n    let end = Math.min(start + maxChars, trimmed.length);\n    // Prefer breaking on whitespace, but only if it doesn't shrink the chunk too much.\n    if (end < trimmed.length) {\n      const lastSpace = trimmed.lastIndexOf(\" \", end);\n      if (lastSpace > start + maxChars / 2) end = lastSpace;\n    }\n    chunks.push(trimmed.slice(start, end).trim());\n    if (end >= trimmed.length) break;\n    start = end - overlap;\n  }\n  return chunks.filter((c) => c.length > 0);\n}\n"],"mappings":";;;;;;;;AAgBA,SAAgB,UAAU,MAAc,UAAwB,CAAC,GAAa;CAC5E,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,UAAU,KAAK,IAAI,QAAQ,WAAW,KAAK,KAAK,MAAM,WAAW,CAAC,CAAC;CACzE,MAAM,UAAU,KAAK,KAAK;CAC1B,IAAI,CAAC,SAAS,OAAO,CAAC;CACtB,IAAI,QAAQ,UAAU,UAAU,OAAO,CAAC,OAAO;CAE/C,MAAM,SAAmB,CAAC;CAC1B,IAAI,QAAQ;CACZ,OAAO,QAAQ,QAAQ,QAAQ;EAC7B,IAAI,MAAM,KAAK,IAAI,QAAQ,UAAU,QAAQ,MAAM;EAEnD,IAAI,MAAM,QAAQ,QAAQ;GACxB,MAAM,YAAY,QAAQ,YAAY,KAAK,GAAG;GAC9C,IAAI,YAAY,QAAQ,WAAW,GAAG,MAAM;EAC9C;EACA,OAAO,KAAK,QAAQ,MAAM,OAAO,GAAG,EAAE,KAAK,CAAC;EAC5C,IAAI,OAAO,QAAQ,QAAQ;EAC3B,QAAQ,MAAM;CAChB;CACA,OAAO,OAAO,QAAQ,MAAM,EAAE,SAAS,CAAC;AAC1C"}

package/dist/chunk-BhUZmQg5.js

@@ -0,0 +1,32 @@
+//#region src/core/chunk.ts
+/**
+* Split long text into overlapping chunks for embedding/indexing. Each chunk is
+* at most `maxChars`; chunks overlap by `overlap` characters so a query that
+* straddles a boundary still matches. Splits prefer the nearest whitespace
+* before the limit to avoid cutting words mid-token. Short text returns a single
+* trimmed chunk; empty/whitespace-only text returns no chunks.
+*/
+function chunkText(text, options = {}) {
+	const maxChars = options.maxChars ?? 1500;
+	const overlap = Math.min(options.overlap ?? 150, Math.floor(maxChars / 2));
+	const trimmed = text.trim();
+	if (!trimmed) return [];
+	if (trimmed.length <= maxChars) return [trimmed];
+	const chunks = [];
+	let start = 0;
+	while (start < trimmed.length) {
+		let end = Math.min(start + maxChars, trimmed.length);
+		if (end < trimmed.length) {
+			const lastSpace = trimmed.lastIndexOf(" ", end);
+			if (lastSpace > start + maxChars / 2) end = lastSpace;
+		}
+		chunks.push(trimmed.slice(start, end).trim());
+		if (end >= trimmed.length) break;
+		start = end - overlap;
+	}
+	return chunks.filter((c) => c.length > 0);
+}
+//#endregion
+export { chunkText as t };
+
+//# sourceMappingURL=chunk-BhUZmQg5.js.map

package/dist/chunk-BhUZmQg5.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"chunk-BhUZmQg5.js","names":[],"sources":["../src/core/chunk.ts"],"sourcesContent":["// src/core/chunk.ts\n\nexport interface ChunkOptions {\n  /** Target maximum characters per chunk. */\n  maxChars?: number;\n  /** Characters of overlap carried from the end of one chunk into the next. */\n  overlap?: number;\n}\n\n/**\n * Split long text into overlapping chunks for embedding/indexing. Each chunk is\n * at most `maxChars`; chunks overlap by `overlap` characters so a query that\n * straddles a boundary still matches. Splits prefer the nearest whitespace\n * before the limit to avoid cutting words mid-token. Short text returns a single\n * trimmed chunk; empty/whitespace-only text returns no chunks.\n */\nexport function chunkText(text: string, options: ChunkOptions = {}): string[] {\n  const maxChars = options.maxChars ?? 1500;\n  const overlap = Math.min(options.overlap ?? 150, Math.floor(maxChars / 2));\n  const trimmed = text.trim();\n  if (!trimmed) return [];\n  if (trimmed.length <= maxChars) return [trimmed];\n\n  const chunks: string[] = [];\n  let start = 0;\n  while (start < trimmed.length) {\n    let end = Math.min(start + maxChars, trimmed.length);\n    // Prefer breaking on whitespace, but only if it doesn't shrink the chunk too much.\n    if (end < trimmed.length) {\n      const lastSpace = trimmed.lastIndexOf(\" \", end);\n      if (lastSpace > start + maxChars / 2) end = lastSpace;\n    }\n    chunks.push(trimmed.slice(start, end).trim());\n    if (end >= trimmed.length) break;\n    start = end - overlap;\n  }\n  return chunks.filter((c) => c.length > 0);\n}\n"],"mappings":";;;;;;;;AAgBA,SAAgB,UAAU,MAAc,UAAwB,CAAC,GAAa;CAC5E,MAAM,WAAW,QAAQ,YAAY;CACrC,MAAM,UAAU,KAAK,IAAI,QAAQ,WAAW,KAAK,KAAK,MAAM,WAAW,CAAC,CAAC;CACzE,MAAM,UAAU,KAAK,KAAK;CAC1B,IAAI,CAAC,SAAS,OAAO,CAAC;CACtB,IAAI,QAAQ,UAAU,UAAU,OAAO,CAAC,OAAO;CAE/C,MAAM,SAAmB,CAAC;CAC1B,IAAI,QAAQ;CACZ,OAAO,QAAQ,QAAQ,QAAQ;EAC7B,IAAI,MAAM,KAAK,IAAI,QAAQ,UAAU,QAAQ,MAAM;EAEnD,IAAI,MAAM,QAAQ,QAAQ;GACxB,MAAM,YAAY,QAAQ,YAAY,KAAK,GAAG;GAC9C,IAAI,YAAY,QAAQ,WAAW,GAAG,MAAM;EAC9C;EACA,OAAO,KAAK,QAAQ,MAAM,OAAO,GAAG,EAAE,KAAK,CAAC;EAC5C,IAAI,OAAO,QAAQ,QAAQ;EAC3B,QAAQ,MAAM;CAChB;CACA,OAAO,OAAO,QAAQ,MAAM,EAAE,SAAS,CAAC;AAC1C"}

package/dist/chunk-ChC83jai.js

@@ -0,0 +1,2 @@
+import { t as chunkText } from "./chunk-BhUZmQg5.js";
+export { chunkText };