@dsh-libing/knowledge-base 0.1.3 → 0.1.4

package/lib/config.js

@@ -10,4 +10,12 @@ export function resolveConfig(raw = {}) {
     return settingsSchema.parse({ ...raw, dataRoot });
 }
 const defaults = resolveConfig();
-export const Config = Schema.object(Object.fromEntries(Object.entries(defaults).filter(([key]) => !['skipImages', 'acceptPartial'].includes(key)).map(([key, value]) => [key, typeof value === 'number' ? Schema.number().min(1).default(value) : typeof value === 'string' ? Schema.string().default(value) : Schema.boolean().default(value)])));
+export const Config = Schema.object(Object.fromEntries(Object.entries(defaults).filter(([key]) => !['skipImages', 'acceptPartial'].includes(key)).map(([key, value]) => [key, (() => {
+        if (typeof value === 'number') {
+            return Schema.number().min(1).default(value);
+        }
+        if (typeof value === 'string') {
+            return Schema.string().default(value);
+        }
+        return Schema.boolean().default(value);
+    })()])));

package/lib/document/convert.js

@@ -12,7 +12,15 @@ import { mkdtemp, readFile, rm } from 'node:fs/promises';
 import { tmpdir } from 'node:os';
 const parser = new XMLParser({ ignoreAttributes: false, removeNSPrefix: true, parseTagValue: false, trimValues: false });
 const xmlSchema = z.lazy(() => z.union([z.string(), z.number(), z.boolean(), z.null(), z.array(xmlSchema), z.record(z.string(), xmlSchema)]));
-const list = (value) => value === undefined ? [] : Array.isArray(value) ? value : [value];
+const list = (value) => {
+    if (value === undefined) {
+        return [];
+    }
+    if (Array.isArray(value)) {
+        return value;
+    }
+    return [value];
+};
 const object = (value) => value !== null && typeof value === 'object' && !Array.isArray(value) ? value : {};
 function child(value, ...keys) {
     for (const key of keys)
@@ -43,8 +51,13 @@ function wordText(value) {
 }
 const normalize = (value) => value.replace(/\s/gu, '');
 function audit(files) {
-    const doc = parseXml(required(files['word/document.xml'])), styles = files['word/styles.xml'] ? parseXml(files['word/styles.xml']) : {};
-    const warnings = [], styleLevels = {}, headingLevels = {}, outlineLevels = {}, anchors = new Set();
+    const doc = parseXml(required(files['word/document.xml']));
+    const styles = files['word/styles.xml'] ? parseXml(files['word/styles.xml']) : {};
+    const warnings = [];
+    const styleLevels = {};
+    const headingLevels = {};
+    const outlineLevels = {};
+    const anchors = new Set();
     const styleRows = list(child(styles, 'styles', 'style'));
     const resolveLevel = (styleId, seen = new Set()) => {
         if (seen.has(styleId) || !styleId)
@@ -53,19 +66,33 @@ function audit(files) {
         const row = styleRows.find(v => xmlString(child(v, '@_styleId')) === styleId);
         if (!row)
             return;
-        const explicit = child(row, 'pPr', 'outlineLvl', '@_val'), name = xmlString(child(row, 'name', '@_val')) || styleId, common = /^(?:heading|标题)\s*([1-9])$/iu.exec(name);
-        return explicit !== undefined && Number(explicit) < 9 ? Number(explicit) + 1 : common ? Number(common[1]) : resolveLevel(xmlString(child(row, 'basedOn', '@_val')), seen);
+        const explicit = child(row, 'pPr', 'outlineLvl', '@_val');
+        const name = xmlString(child(row, 'name', '@_val')) || styleId;
+        const common = /^(?:heading|标题)\s*([1-9])$/iu.exec(name);
+        if (explicit !== undefined && Number(explicit) < 9) {
+            return Number(explicit) + 1;
+        }
+        if (common) {
+            return Number(common[1]);
+        }
+        return resolveLevel(xmlString(child(row, 'basedOn', '@_val')), seen);
     };
     for (const row of styleRows) {
-        const styleId = xmlString(child(row, '@_styleId')), level = resolveLevel(styleId);
+        const styleId = xmlString(child(row, '@_styleId'));
+        const level = resolveLevel(styleId);
         if (level)
             styleLevels[styleId] = level;
     }
-    let expectedFigures = 0, expectedTables = 0, textBoxes = 0, footnoteRefs = 0;
+    let expectedFigures = 0;
+    let expectedTables = 0;
+    let textBoxes = 0;
+    let footnoteRefs = 0;
     const expectedHeadings = [];
     visit(doc, (key, value) => {
         if (key === 'p') {
-            const outline = child(value, 'pPr', 'outlineLvl', '@_val'), style = xmlString(child(value, 'pPr', 'pStyle', '@_val')), level = outline !== undefined && Number(outline) < 9 ? Number(outline) + 1 : styleLevels[style];
+            const outline = child(value, 'pPr', 'outlineLvl', '@_val');
+            const style = xmlString(child(value, 'pPr', 'pStyle', '@_val'));
+            const level = outline !== undefined && Number(outline) < 9 ? Number(outline) + 1 : styleLevels[style];
             if (outline !== undefined && Number(outline) < 9)
                 outlineLevels[normalize(wordText(value))] = Number(outline) + 1;
             if (level && !/^TOC/iu.test(style)) {
@@ -119,7 +146,8 @@ function safeLink(value) { return /^(?:https?:|mailto:|#)/i.test(value || '') ?
 function inline(node, internalRefs) {
     if ('value' in node)
         return node.value.replace(/([\\`*[\]<>])/g, '\\$1');
-    const t = tag(node), body = children(node).map(n => inline(n, internalRefs)).join('');
+    const t = tag(node);
+    const body = children(node).map(n => inline(n, internalRefs)).join('');
     if (['script', 'style', 'iframe', 'object'].includes(t))
         return '';
     if (t === 'img')
@@ -159,8 +187,10 @@ function tableHtml(rows) {
 }
 export async function convertDocx(buffer, destination, config, signal) {
     checkAbort(signal);
-    const files = unpackDocx(buffer, config), originalAudit = audit(files);
-    const assets = [], warnings = [...originalAudit.warnings];
+    const files = unpackDocx(buffer, config);
+    const originalAudit = audit(files);
+    const assets = [];
+    const warnings = [...originalAudit.warnings];
     const mimeExtension = { 'image/png': 'png', 'image/jpeg': 'jpg', 'image/gif': 'gif', 'image/webp': 'webp', 'image/bmp': 'bmp', 'image/tiff': 'tif', 'image/x-emf': 'emf', 'image/x-wmf': 'wmf', 'image/svg+xml': 'svg' };
     const addAsset = async (data, mime, origin) => {
         const assetId = `a-${hash(data).slice(0, 20)}`;
@@ -174,7 +204,8 @@ export async function convertDocx(buffer, destination, config, signal) {
     };
     const rawMedia = Object.entries(files).filter(([name]) => name.startsWith('word/media/'));
     for (const [name, bytes] of rawMedia) {
-        const ext = (name.split('.').at(-1) || '').toLowerCase(), mime = Object.entries(mimeExtension).find(([, x]) => x === ext || ext === 'jpeg' && x === 'jpg')?.[0] || 'application/octet-stream';
+        const ext = (name.split('.').at(-1) || '').toLowerCase();
+        const mime = Object.entries(mimeExtension).find(([, x]) => x === ext || ext === 'jpeg' && x === 'jpg')?.[0] || 'application/octet-stream';
         await addAsset(Buffer.from(bytes), mime, name);
     }
     const result = await mammoth.convertToHtml({ buffer }, {
@@ -208,11 +239,14 @@ export async function convertDocx(buffer, destination, config, signal) {
         // Unmapped styles lose formatting, not paragraph text; heading coverage is audited separately.
         blocking: !/^Unrecognised (?:paragraph|run) style:/i.test(m.message) && /image|unrecognised|unsupported/i.test(m.message),
     })));
-    const fragment = parseFragment(result.value), blocks = [], internalRefs = [];
+    const fragment = parseFragment(result.value);
+    const blocks = [];
+    const internalRefs = [];
     let figureCount = 0;
     const add = (value) => { blocks.push(makeBlock({ blockId: `b-${String(blocks.length + 1).padStart(6, '0')}`, ...value })); return required(blocks.at(-1)); };
     const addInline = (node, type = 'paragraph', prefix = '') => {
-        const body = inline(node, internalRefs), pieces = body.split(String.fromCharCode(1)).filter(Boolean);
+        const body = inline(node, internalRefs);
+        const pieces = body.split(String.fromCharCode(1)).filter(Boolean);
         for (const piece of pieces) {
             if (piece.startsWith('asset:')) {
                 figureCount++;
@@ -227,7 +261,8 @@ export async function convertDocx(buffer, destination, config, signal) {
         if (/^h[1-9]$/.test(t))
             add({ type: 'heading', level: Number(t[1]), text: htmlText(node).trim(), anchor: attr(node, 'id') || null });
         else if (t === 'table') {
-            const rows = tableCells(node), table = add({ type: 'table', rows, text: rows.map(row => row.map(c => c.text).join(' | ')).join('\n'), markdown: tableHtml(rows) });
+            const rows = tableCells(node);
+            const table = add({ type: 'table', rows, text: rows.map(row => row.map(c => c.text).join(' | ')).join('\n'), markdown: tableHtml(rows) });
             const scan = (n) => {
                 if (tag(n) === 'img') {
                     const assetId = (attr(n, 'src') || '').replace(/^asset:/, '');
@@ -241,7 +276,8 @@ export async function convertDocx(buffer, destination, config, signal) {
             scan(node);
         }
         else if (t === 'p') {
-            const t = htmlText(node).trim(), previous = blocks.at(-1);
+            const t = htmlText(node).trim();
+            const previous = blocks.at(-1);
             if (/^(?:图|Figure)\s*[A-Za-z]?\d[\d.\-–]*/i.test(t) && previous?.type === 'figure') {
                 previous.caption = t;
                 add({ type: 'caption', text: t, figureRef: previous.occurrenceId });

package/lib/document/pdf-layout.js

@@ -19,7 +19,8 @@ export function lines(spans) {
     }
     return rows.sort((a, b) => a.y - b.y).map(r => {
         const items = r.spans.sort((a, b) => readingBox(a)[0] - readingBox(b)[0]);
-        let text = '', last;
+        let text = '';
+        let last;
         for (const s of items) {
             if (last && readingBox(s)[0] - readingBox(last)[2] > Math.max(1, s.fontSize * .18) && !/\s$/.test(text))
                 text += ' ';
@@ -83,16 +84,23 @@ function geometricTables(page, spans, occupied) {
             group.push(required(rows[++i]));
         if (group.length < 2)
             continue;
-        const ids = new Set(), cells = group.map(r => r.cells.map(b => { const ss = spans.filter(s => { const x = (s.bbox[0] + s.bbox[2]) / 2, y = (s.bbox[1] + s.bbox[3]) / 2; return x > b[0] && x < b[2] && y > b[1] && y < b[3]; }); ss.forEach(s => ids.add(s.id)); return { text: textOf(ss), spanIds: ss.map(s => s.id), bbox: b, colspan: 1, rowspan: 1 }; }));
+        const ids = new Set();
+        const cells = group.map(r => r.cells.map(b => { const ss = spans.filter(s => { const x = (s.bbox[0] + s.bbox[2]) / 2; const y = (s.bbox[1] + s.bbox[3]) / 2; return x > b[0] && x < b[2] && y > b[1] && y < b[3]; }); ss.forEach(s => ids.add(s.id)); return { text: textOf(ss), spanIds: ss.map(s => s.id), bbox: b, colspan: 1, rowspan: 1 }; }));
         if (cells.some(r => r.every(c => !c.text)))
             continue;
-        const bbox = required(union(group.flatMap(r => r.cells))), identified = /field|offset|bits|name|description|字段|偏移|名称|参数|描述|位范围/i.test(required(cells[0]).map(c => c.text).join(' '));
+        const bbox = required(union(group.flatMap(r => r.cells)));
+        const identified = /field|offset|bits|name|description|字段|偏移|名称|参数|描述|位范围/i.test(required(cells[0]).map(c => c.text).join(' '));
         tables.push({ rows: cells, bbox, ids, identified });
     }
     return tables;
 }
 export function nativeLayout(page, info) {
-    const byMarked = new Map(), used = new Set(), blocks = [], figures = [], warnings = (page.warnings || []).map(message => ({ code: 'pdf-engine-warning', page: page.page, message: `PDF 第 ${page.page} 页解析器警告：${message}`, blocking: true })), excluded = page.spans.filter(s => s.artifact).map(s => ({ kind: 'tagged-artifact', text: s.text, bbox: s.bbox }));
+    const byMarked = new Map();
+    const used = new Set();
+    const blocks = [];
+    const figures = [];
+    const warnings = (page.warnings || []).map(message => ({ code: 'pdf-engine-warning', page: page.page, message: `PDF 第 ${page.page} 页解析器警告：${message}`, blocking: true }));
+    const excluded = page.spans.filter(s => s.artifact).map(s => ({ kind: 'tagged-artifact', text: s.text, bbox: s.bbox }));
     const textSpans = page.spans.filter(s => !s.artifact);
     for (const s of textSpans) {
         if (!byMarked.has(s.mcid))
@@ -128,7 +136,8 @@ export function nativeLayout(page, info) {
             if (!rows.length)
                 return;
             const irregular = rows.some(r => r.length !== required(rows[0]).length);
-            const textBox = required(union(spans.map(s => s.bbox))), nearby = page.drawings.filter(d => d.kind === 'path' && d.bbox[1] >= textBox[1] - 20 && d.bbox[3] <= textBox[3] + 20 && intersects(d.bbox, boundedBox(textBox, page, 10))).map(d => d.bbox);
+            const textBox = required(union(spans.map(s => s.bbox)));
+            const nearby = page.drawings.filter(d => d.kind === 'path' && d.bbox[1] >= textBox[1] - 20 && d.bbox[3] <= textBox[3] + 20 && intersects(d.bbox, boundedBox(textBox, page, 10))).map(d => d.bbox);
             const tableBox = boundedBox(required(union([textBox, ...nearby])), page, 3);
             add('table', spans, { source: source(tableBox), rows, markdown: tableMarkdown(rows), needsReview: irregular, caption: '', layoutBasis: 'structure-tags', ...textBox[3] > page.height * .85 && !closedTableBottom(page, textBox, tableBox) ? { openEnd: true } : {} });
             if (irregular)
@@ -136,7 +145,10 @@ export function nativeLayout(page, info) {
             return;
         }
         if (node.role === 'Figure') {
-            const marked = ids(node), spans = get(node), boxes = [...page.drawings.filter(d => d.mcid !== null && marked.includes(d.mcid)).map(d => d.bbox), ...spans.map(s => s.bbox)], bbox = union(boxes);
+            const marked = ids(node);
+            const spans = get(node);
+            const boxes = [...page.drawings.filter(d => d.mcid !== null && marked.includes(d.mcid)).map(d => d.bbox), ...spans.map(s => s.bbox)];
+            const bbox = union(boxes);
             if (!bbox)
                 return;
             consume(spans);
@@ -148,7 +160,15 @@ export function nativeLayout(page, info) {
             return;
         }
         if (['Code', 'Formula', 'Note'].includes(node.role || '')) {
-            add(node.role === 'Code' ? 'code' : node.role === 'Formula' ? 'formula' : 'note', get(node));
+            add((() => {
+                if (node.role === 'Code') {
+                    return 'code';
+                }
+                if (node.role === 'Formula') {
+                    return 'formula';
+                }
+                return 'note';
+            })(), get(node));
             return;
         }
         if (node.role === 'P') {
@@ -194,8 +214,21 @@ export function nativeLayout(page, info) {
         const match = info.outline.find(o => {
             if (o.page !== page.page || normalize(o.title) !== normalize(b.text))
                 return false;
-            const candidates = blocks.filter(c => normalize(c.text) === normalize(o.title)), t = page.viewportTransform, axis = page.rotation % 180 ? 0 : 1;
-            const target = o.y !== null && Number.isFinite(o.y) && t ? (axis === 0 ? required(t[2]) * o.y + required(t[4]) : required(t[3]) * o.y + required(t[5])) : null;
+            const candidates = blocks.filter(c => normalize(c.text) === normalize(o.title));
+            const t = page.viewportTransform;
+            const axis = page.rotation % 180 ? 0 : 1;
+            let target;
+            if (o.y !== null && Number.isFinite(o.y) && t) {
+                if (axis === 0) {
+                    target = required(t[2]) * o.y + required(t[4]);
+                }
+                else {
+                    target = required(t[3]) * o.y + required(t[5]);
+                }
+            }
+            else {
+                target = null;
+            }
             const score = (c) => target === null ? (c.type === 'heading' ? 0 : 100) + c.source.bbox[1] / 1000 : Math.abs(c.source.bbox[axis] - target);
             return candidates.sort((a, b) => score(a) - score(b))[0] === b && (target === null || score(b) < 100);
         });
@@ -225,7 +258,9 @@ export function nativeLayout(page, info) {
     blocks.sort((a, b) => a.source.bbox[1] - b.source.bbox[1] || a.source.bbox[0] - b.source.bbox[0]);
     const coverage = [...[...figures, ...blocks.filter(b => b.type === 'table')].map(b => boundedBox(b.source.bbox, page, 15)), ...excluded.map(e => boundedBox(e.bbox, page, 2))];
     const unexplained = page.drawings.filter(d => d.bbox[2] - d.bbox[0] > 25 && d.bbox[3] - d.bbox[1] > 12 && d.bbox[1] > 65 && d.bbox[3] < page.height - 40 && !coverage.some(b => intersects(b, d.bbox)));
-    const raw = textSpans.map(s => s.text).join(''), garbled = textSpans.filter(s => (s.text.includes('\ufffd') || s.text.includes('\0'))), ocr = (raw.length < 30 && page.drawings.length > 0) || garbled.length > textSpans.length * .3;
+    const raw = textSpans.map(s => s.text).join('');
+    const garbled = textSpans.filter(s => (s.text.includes('\ufffd') || s.text.includes('\0')));
+    const ocr = (raw.length < 30 && page.drawings.length > 0) || garbled.length > textSpans.length * .3;
     if (garbled.length)
         warnings.push({ code: 'pdf-garbled-text', page: page.page, message: `PDF 第 ${page.page} 页存在缺字或乱码，需要核对识别区域`, blocking: true });
     const multiColumn = lines(textSpans.filter(s => !coverage.some(b => intersects(s.bbox, b)))).filter(l => l.spans.some((s, i) => i && s.bbox[0] - required(l.spans[i - 1]).bbox[2] > page.width * .13)).length >= 3;

package/lib/document/pdf-session.js

@@ -24,7 +24,8 @@ export class PdfSession {
         this.signal = signal;
         checkAbort(signal);
         // Production uses the shipped JS worker. Source tests use the same TS through the installed tsx loader.
-        const js = new URL('./pdf-worker.js', import.meta.url), source = !existsSync(fileURLToPath(js));
+        const js = new URL('./pdf-worker.js', import.meta.url);
+        const source = !existsSync(fileURLToPath(js));
         this.worker = new Worker(source ? new URL('./pdf-worker.ts', import.meta.url) : js, { workerData: { file, config: { maxInputBytes: config.maxInputBytes, pdfMaxPages: config.pdfMaxPages, pdfMaxPagePixels: config.pdfMaxPagePixels } }, resourceLimits: { maxOldGenerationSizeMb: 512 }, execArgv: source ? ['--import', 'tsx/esm'] : [] });
         const open = deferred();
         this.ready = open.promise;
@@ -85,7 +86,8 @@ export class PdfSession {
             throw this.error;
         if (this.closed)
             throw new Error('PDF 解析进程已关闭');
-        const id = ++this.seq, value = await new Promise((resolve, reject) => { const timer = setTimeout(() => this.fail(new Error(`PDF 第 ${page} 页处理超时`)), this.config.pdfPageTimeoutMs); this.pending.set(id, { resolve, reject, timer }); this.worker.postMessage({ id, command, page, bbox }); });
+        const id = ++this.seq;
+        const value = await new Promise((resolve, reject) => { const timer = setTimeout(() => this.fail(new Error(`PDF 第 ${page} 页处理超时`)), this.config.pdfPageTimeoutMs); this.pending.set(id, { resolve, reject, timer }); this.worker.postMessage({ id, command, page, bbox }); });
         return command === 'page' ? nativePageSchema.parse(value) : renderSchema.parse(value);
     }
     close() {
@@ -114,9 +116,13 @@ export function pageRange(start, end, count) {
 export async function inspectPdf(file, config, signal) {
     const session = new PdfSession(file, config, signal);
     try {
-        const info = await session.ready, n = info.pageCount, samples = [...new Set([1, Math.min(3, n), Math.ceil(n / 3), Math.ceil(n * 2 / 3), n])], pages = [];
+        const info = await session.ready;
+        const n = info.pageCount;
+        const samples = [...new Set([1, Math.min(3, n), Math.ceil(n / 3), Math.ceil(n * 2 / 3), n])];
+        const pages = [];
         for (const page of samples) {
-            const value = await session.call('page', page), text = value.spans.filter(s => !s.artifact).map(s => s.text).join('');
+            const value = await session.call('page', page);
+            const text = value.spans.filter(s => !s.artifact).map(s => s.text).join('');
             pages.push({ page, chars: Array.from(text).length, hasGraphics: value.drawings.length > 0, possibleOcr: text.length < 30 && value.drawings.length > 0, hasStructure: Boolean(value.tree) });
         }
         return { sourceFormat: 'pdf', pageCount: n, outlineCount: info.outline.length, samples: pages, possibleOcr: pages.some(p => p.possibleOcr), note: '仅抽样预检；正式构建逐页核对，图表及扫描识别共用模型预算。' };

package/lib/document/pdf-vision.js

@@ -16,7 +16,9 @@ export async function restoreLayout(page, layout, file, job, models, signal) {
         throw new Error('此页需要版面识别/OCR，请配置图片模型并启用图片理解');
     if (!job.visionCapability)
         await models.preflight(job, signal);
-    const spans = page.spans.filter(s => !s.artifact), ocr = layout.ocr, badIds = new Set(spans.filter(s => (s.text.includes('\ufffd') || s.text.includes('\0'))).map(s => s.id));
+    const spans = page.spans.filter(s => !s.artifact);
+    const ocr = layout.ocr;
+    const badIds = new Set(spans.filter(s => (s.text.includes('\ufffd') || s.text.includes('\0'))).map(s => s.id));
     if (chars(JSON.stringify(spans)) > 90000)
         throw new Error('此页原生文字对象超过版面识别输入上限，请缩小页范围或使用专用解析器');
     const prepared = await prepareImage(models.ctx, file, job.config, signal);
@@ -29,7 +31,8 @@ export async function restoreLayout(page, layout, file, job, models, signal) {
         const value = assignmentSchema.parse(raw);
         if (!Array.isArray(value.blocks) || !value.blocks.length || value.blocks.length > 2000 || !Array.isArray(value.uncertainties) || value.uncertainties.length > 100)
             throw new Error('PDF 版面识别 JSON 无效');
-        const used = new Set(), byId = new Map(spans.map(s => [s.id, s]));
+        const used = new Set();
+        const byId = new Map(spans.map(s => [s.id, s]));
         const take = (row, allowEmpty = false) => {
             if (ocr && (!row.spanIds || Array.isArray(row.spanIds) && !row.spanIds.length)) {
                 if (typeof row.text !== 'string' || chars(row.text) > 30000 || !allowEmpty && !row.text.trim())
@@ -97,7 +100,9 @@ export async function restoreLayout(page, layout, file, job, models, signal) {
             }
             // The assigned strings must lie inside their claimed region. No numeric text rewriting is accepted.
             for (const id of block.spanIds || []) {
-                const s = required(byId.get(id)), x = (s.bbox[0] + s.bbox[2]) / 2, y = (s.bbox[1] + s.bbox[3]) / 2;
+                const s = required(byId.get(id));
+                const x = (s.bbox[0] + s.bbox[2]) / 2;
+                const y = (s.bbox[1] + s.bbox[3]) / 2;
                 if (x < b[0] - 2 || x > b[2] + 2 || y < b[1] - 2 || y > b[3] + 2)
                     throw new Error('PDF 文字与识别区域不匹配');
             }

package/lib/document/pdf-worker.js

@@ -14,11 +14,14 @@ const matrixSchema = z.tuple([z.number(), z.number(), z.number(), z.number(), z.
 const markTag = (value) => z.union([z.string(), z.object({ name: z.string() })]).transform(value => typeof value === 'string' ? value : value.name).parse(value);
 const numeric = (value) => z.array(z.number().finite()).parse(value instanceof Float32Array || value instanceof Float64Array ? Array.from(value) : value);
 // PDF.js uses napi-rs on Node. Correct only subpixel clip skew in this isolated worker's supported canvas prototype.
-const proto = Object.getPrototypeOf(createCanvas(1, 1).getContext('2d')), nativeClip = proto.clip;
+const proto = Object.getPrototypeOf(createCanvas(1, 1).getContext('2d'));
+const nativeClip = proto.clip;
 proto.clip = function (path, rule) {
-    const original = this.getTransform(), tiny = (value) => value !== 0 && Math.abs(value) < 1e-7;
+    const original = this.getTransform();
+    const tiny = (value) => value !== 0 && Math.abs(value) < 1e-7;
     if (path instanceof Path2D && (tiny(original.b) || tiny(original.c))) {
-        const matrix = this.getTransform(), transformed = new Path2D();
+        const matrix = this.getTransform();
+        const transformed = new Path2D();
         if (tiny(matrix.b))
             matrix.b = 0;
         if (tiny(matrix.c))
@@ -33,13 +36,25 @@ proto.clip = function (path, rule) {
         }
     }
     else
-        Reflect.apply(nativeClip, this, path === undefined ? [] : rule === undefined ? [path] : [path, rule]);
+        Reflect.apply(nativeClip, this, (() => {
+            if (path === undefined) {
+                return [];
+            }
+            if (rule === undefined) {
+                return [path];
+            }
+            return [path, rule];
+        })());
 };
-const require = createRequire(import.meta.url), base = dirname(require.resolve('pdfjs-dist/package.json')), config = input.config;
-let doc, activePage = 0;
+const require = createRequire(import.meta.url);
+const base = dirname(require.resolve('pdfjs-dist/package.json'));
+const config = input.config;
+let doc;
+let activePage = 0;
 const engineWarnings = new Map();
 console.warn = (...parts) => {
-    const value = parts.map(String).join(' ').replace(/^Warning:\s*/u, '').slice(0, 1000), list = engineWarnings.get(activePage) || [];
+    const value = parts.map(String).join(' ').replace(/^Warning:\s*/u, '').slice(0, 1000);
+    const list = engineWarnings.get(activePage) || [];
     if (list.length < 100 && !list.includes(value))
         list.push(value);
     engineWarnings.set(activePage, list);
@@ -70,7 +85,8 @@ async function info() {
         for (const entry of items || []) {
             if (outline.length >= 10000 || level > 32)
                 throw new Error('PDF 书签层级或数量超过上限');
-            let page = null, y = null;
+            let page = null;
+            let y = null;
             try {
                 const raw = typeof entry.dest === 'string' ? await doc.getDestination(entry.dest) : entry.dest;
                 if (Array.isArray(raw) && raw.length) {
@@ -93,8 +109,11 @@ async function nativePage(n) {
     activePage = n;
     const page = await doc.getPage(n);
     try {
-        const viewport = page.getViewport({ scale: 1 }), content = await page.getTextContent({ includeMarkedContent: true }), tree = prune(await page.getStructTree());
-        const spans = [], marks = [];
+        const viewport = page.getViewport({ scale: 1 });
+        const content = await page.getTextContent({ includeMarkedContent: true });
+        const tree = prune(await page.getStructTree());
+        const spans = [];
+        const marks = [];
         let seq = 0;
         for (const item of content.items) {
             if ('type' in item) {
@@ -106,17 +125,47 @@ async function nativePage(n) {
             }
             if (!item.str.trim())
                 continue;
-            const t = matrixSchema.parse(Util.transform(viewport.transform, item.transform)), h = Math.hypot(t[2], t[3]) || item.height || 10, style = content.styles[item.fontName], a = style?.ascent ?? .8;
-            const a0 = [t[4] + t[2] * a, t[5] + t[3] * a], b0 = [t[4] + t[2] * (a - 1), t[5] + t[3] * (a - 1)];
-            const angle = Math.atan2(t[1], t[0]), w = item.width || h / 2, dx = Math.cos(angle) * w, dy = Math.sin(angle) * w;
-            const points = [a0, b0, [a0[0] + dx, a0[1] + dy], [b0[0] + dx, b0[1] + dy]], bbox = [Math.min(...points.map(p => p[0])), Math.min(...points.map(p => p[1])), Math.max(...points.map(p => p[0])), Math.max(...points.map(p => p[1]))];
-            const rotation = ((page.rotate % 360) + 360) % 360, W = viewport.width, H = viewport.height;
-            const readingBox = rotation === 90 ? [bbox[1], W - bbox[2], bbox[3], W - bbox[0]] : rotation === 180 ? [W - bbox[2], H - bbox[3], W - bbox[0], H - bbox[1]] : rotation === 270 ? [H - bbox[3], bbox[0], H - bbox[1], bbox[2]] : bbox;
+            const t = matrixSchema.parse(Util.transform(viewport.transform, item.transform));
+            const h = Math.hypot(t[2], t[3]) || item.height || 10;
+            const style = content.styles[item.fontName];
+            const a = style?.ascent ?? .8;
+            const a0 = [t[4] + t[2] * a, t[5] + t[3] * a];
+            const b0 = [t[4] + t[2] * (a - 1), t[5] + t[3] * (a - 1)];
+            const angle = Math.atan2(t[1], t[0]);
+            const w = item.width || h / 2;
+            const dx = Math.cos(angle) * w;
+            const dy = Math.sin(angle) * w;
+            const points = [a0, b0, [a0[0] + dx, a0[1] + dy], [b0[0] + dx, b0[1] + dy]];
+            const bbox = [Math.min(...points.map(p => p[0])), Math.min(...points.map(p => p[1])), Math.max(...points.map(p => p[0])), Math.max(...points.map(p => p[1]))];
+            const rotation = ((page.rotate % 360) + 360) % 360;
+            const W = viewport.width;
+            const H = viewport.height;
+            let readingBox;
+            if (rotation === 90) {
+                readingBox = [bbox[1], W - bbox[2], bbox[3], W - bbox[0]];
+            }
+            else {
+                if (rotation === 180) {
+                    readingBox = [W - bbox[2], H - bbox[3], W - bbox[0], H - bbox[1]];
+                }
+                else {
+                    if (rotation === 270) {
+                        readingBox = [H - bbox[3], bbox[0], H - bbox[1], bbox[2]];
+                    }
+                    else {
+                        readingBox = bbox;
+                    }
+                }
+            }
             spans.push({ id: `s${++seq}`, text: item.str, mcid: [...marks].reverse().find(m => m.id)?.id || null, artifact: marks.some(m => m.artifact), bbox, readingBox, fontSize: h, fontName: style?.fontFamily || item.fontName });
             if (spans.length > 100000)
                 throw new Error(`PDF 第 ${n} 页文字对象过多`);
         }
-        const operators = await page.getOperatorList(), drawings = [], stack = [], mc = [], prefix = `p${page.ref?.num || n}R_mc`;
+        const operators = await page.getOperatorList();
+        const drawings = [];
+        const stack = [];
+        const mc = [];
+        const prefix = `p${page.ref?.num || n}R_mc`;
         let matrix = [1, 0, 0, 1, 0, 0];
         const add = (b, kind, shape = null) => {
             const bounds = box(matrixSchema.parse(Util.transform(viewport.transform, matrix)), b);
@@ -126,7 +175,8 @@ async function nativePage(n) {
             drawings.push({ bbox: bounds, kind, shape, mcid: id === undefined ? null : prefix + id });
         };
         for (let i = 0; i < operators.fnArray.length; i++) {
-            const op = required(operators.fnArray[i]), args = operators.argsArray[i] || [];
+            const op = required(operators.fnArray[i]);
+            const args = operators.argsArray[i] || [];
             if (op === OPS.save)
                 stack.push([...matrix]);
             else if (op === OPS.restore)
@@ -147,7 +197,8 @@ async function nativePage(n) {
             else if (op === OPS.endMarkedContent)
                 mc.pop();
             else if (op === OPS.constructPath && [OPS.stroke, OPS.fill, OPS.eoFill, OPS.fillStroke, OPS.eoFillStroke].includes(z.number().parse(args[0])) && args[2]) {
-                const paths = Array.isArray(args[1]) ? args[1] : [], path = paths[0] === undefined ? [] : numeric(paths[0]);
+                const paths = Array.isArray(args[1]) ? args[1] : [];
+                const path = paths[0] === undefined ? [] : numeric(paths[0]);
                 add(bboxSchema.parse(numeric(args[2])), 'path', path.length === 13 && path[0] === 0 && path[3] === 1 && path[6] === 1 && path[9] === 1 && path[12] === 4 ? 'rect' : null);
             }
             else if ([OPS.paintImageXObject, OPS.paintInlineImageXObject, OPS.paintImageMaskXObject].includes(op))
@@ -163,13 +214,17 @@ async function nativePage(n) {
 }
 async function render(n, bbox) {
     activePage = n;
-    const page = await doc.getPage(n), v = page.getViewport({ scale: 1 }), b = bbox || [0, 0, v.width, v.height];
+    const page = await doc.getPage(n);
+    const v = page.getViewport({ scale: 1 });
+    const b = bbox || [0, 0, v.width, v.height];
     if (b[0] < 0 || b[1] < 0 || b[2] > v.width + .1 || b[3] > v.height + .1 || b[2] <= b[0] || b[3] <= b[1])
         throw new Error('PDF 裁切区域无效');
     const scale = Math.min(2, Math.sqrt(config.pdfMaxPagePixels / (v.width * v.height)));
     if (scale < .5)
         throw new Error(`PDF 第 ${n} 页页面尺寸过大`);
-    const viewport = page.getViewport({ scale }), canvas = createCanvas(Math.ceil(viewport.width), Math.ceil(viewport.height)), crop = createCanvas(Math.max(1, Math.ceil((b[2] - b[0]) * scale)), Math.max(1, Math.ceil((b[3] - b[1]) * scale)));
+    const viewport = page.getViewport({ scale });
+    const canvas = createCanvas(Math.ceil(viewport.width), Math.ceil(viewport.height));
+    const crop = createCanvas(Math.max(1, Math.ceil((b[2] - b[0]) * scale)), Math.max(1, Math.ceil((b[3] - b[1]) * scale)));
     try {
         // PDF.js publishes DOM-only render types for its supported Node napi-rs canvas. This bridge uses that same tested Canvas2D implementation.
         await page.render({ canvas: null, canvasContext: canvas.getContext('2d'), viewport, background: 'rgb(255,255,255)' }).promise;

package/lib/document/pdf.js

@@ -10,10 +10,14 @@ import { blockMarkdown } from "./convert.js";
 import { BudgetError } from "../models.js";
 export async function convertPdf(file, destination, job, models, signal) {
     let session = new PdfSession(file, job.config, signal);
-    const assets = new Map(), pages = [], blocks = [], warnings = [];
+    const assets = new Map();
+    const pages = [];
+    const blocks = [];
+    const warnings = [];
     job.parserVersion = PDF_PARSER_VERSION;
     const addAsset = async (data, origin, source) => {
-        const assetId = `a-${hash(data).slice(0, 20)}`, asset = { assetId, mime: 'image/png', path: `assets/${assetId}.png`, bytes: data.length, origin, source };
+        const assetId = `a-${hash(data).slice(0, 20)}`;
+        const asset = { assetId, mime: 'image/png', path: `assets/${assetId}.png`, bytes: data.length, origin, source };
         if (!assets.has(assetId)) {
             await atomicWrite(join(destination, asset.path), Buffer.from(data));
             assets.set(assetId, asset);
@@ -21,7 +25,8 @@ export async function convertPdf(file, destination, job, models, signal) {
         return asset;
     };
     try {
-        const info = await session.ready, range = pageRange(job.pageStart, job.pageEnd, info.pageCount);
+        const info = await session.ready;
+        const range = pageRange(job.pageStart, job.pageEnd, info.pageCount);
         warnings.push(...(info.warnings || []).map(message => ({ code: 'pdf-engine-warning', message: `PDF 打开警告：${message}`, blocking: true })));
         job.pdf = { pageCount: info.pageCount, selectedPages: range, outlineCount: info.outline.length };
         job.progress = { ...job.progress, totalPages: range.count, parsedPages: 0, ocrPages: 0, tables: 0 };
@@ -34,7 +39,9 @@ export async function convertPdf(file, destination, job, models, signal) {
             const receipt = join(destination, 'pdf-pages', `${n}-${key}.json`);
             let result = await readJson(receipt, pageReceiptSchema, null);
             if (!result) {
-                let page, layout, previewAsset;
+                let page;
+                let layout;
+                let previewAsset;
                 try {
                     if (session.closed) {
                         session = new PdfSession(file, job.config, signal);
@@ -94,7 +101,15 @@ export async function convertPdf(file, destination, job, models, signal) {
                         }
                         return { ...value, localIndex: i };
                     });
-                    result = { blocks: pageBlocks, assets: localAssets, warnings: layout.warnings, audit: { page: n, width: page.width, height: page.height, pageLabel: info.pageLabels?.[n - 1] || null, status: layout.warnings.some(w => w.blocking) ? 'needs-review' : layout.navigation ? 'navigation' : 'parsed', ocr: layout.ocr, spanCount: layout.spanCount, artifactCount: layout.artifactCount, excluded: layout.excluded, tables: pageBlocks.filter(b => b.type === 'table').length, figures: pageBlocks.filter(b => b.type === 'figure').length, headings: pageBlocks.filter(b => b.type === 'heading').map(b => ({ text: b.text, basis: b.headingBasis, level: b.level })), previewAssetId: previewAsset?.assetId || null } };
+                    result = { blocks: pageBlocks, assets: localAssets, warnings: layout.warnings, audit: { page: n, width: page.width, height: page.height, pageLabel: info.pageLabels?.[n - 1] || null, status: (() => {
+                                if (layout.warnings.some(w => w.blocking)) {
+                                    return 'needs-review';
+                                }
+                                if (layout.navigation) {
+                                    return 'navigation';
+                                }
+                                return 'parsed';
+                            })(), ocr: layout.ocr, spanCount: layout.spanCount, artifactCount: layout.artifactCount, excluded: layout.excluded, tables: pageBlocks.filter(b => b.type === 'table').length, figures: pageBlocks.filter(b => b.type === 'figure').length, headings: pageBlocks.filter(b => b.type === 'heading').map(b => ({ text: b.text, basis: b.headingBasis, level: b.level })), previewAssetId: previewAsset?.assetId || null } };
                 }
                 catch (error) {
                     if (error instanceof BudgetError || signal?.aborted)
@@ -148,10 +163,23 @@ export async function convertPdf(file, destination, job, models, signal) {
         if (pages.length >= 3) {
             const repeated = new Map();
             for (const b of blocks) {
-                const box = b.source?.bbox, p = pages.find(p => p.page === b.source?.page);
+                const box = b.source?.bbox;
+                const p = pages.find(p => p.page === b.source?.page);
                 if (b.type !== 'paragraph' || !box || b.text.length > 160)
                     continue;
-                const key = normalize(b.text).replace(/\d+/g, '#'), edge = box[1] < 40 ? 'top' : box[1] > (p?.height || 842) - 45 ? 'bottom' : null;
+                const key = normalize(b.text).replace(/\d+/g, '#');
+                let edge;
+                if (box[1] < 40) {
+                    edge = 'top';
+                }
+                else {
+                    if (box[1] > (p?.height || 842) - 45) {
+                        edge = 'bottom';
+                    }
+                    else {
+                        edge = null;
+                    }
+                }
                 if (edge) {
                     const k = edge + ':' + key;
                     repeated.set(k, [...(repeated.get(k) || []), b]);
@@ -179,11 +207,13 @@ export async function convertPdf(file, destination, job, models, signal) {
                     repeated.push(item);
                 previous--;
             }
-            const a = blocks[previous], last = a?.source?.locations?.at(-1) || a?.source;
+            const a = blocks[previous];
+            const last = a?.source?.locations?.at(-1) || a?.source;
             const tableId = (t) => /^(?:表|Table)\s*([A-Z\d][\w.\-–]*)/iu.exec(t.caption)?.[1];
             if (!a || a.type !== 'table' || !a.source || !last?.bbox || !a.rows[0] || !b.rows[0])
                 continue;
-            const height = pages.find(p => p.page === last.page)?.height || 842, nextHeight = pages.find(p => p.page === b.source?.page)?.height || 842;
+            const height = pages.find(p => p.page === last.page)?.height || 842;
+            const nextHeight = pages.find(p => p.page === b.source?.page)?.height || 842;
             if (b.source.page === last.page + 1 && tableId(a) && tableId(a) === tableId(b) && last.bbox[3] > height * .8 && b.source.bbox[1] < nextHeight * .25 && !a.needsReview && !b.needsReview && JSON.stringify(a.rows[0].map(c => normalize(c.text))) === JSON.stringify(b.rows[0].map(c => normalize(c.text))) && a.rows[0].length === b.rows[0].length && [...a.rows, ...b.rows].every(r => r.length === required(a.rows[0]).length && r.every(c => c.rowspan === 1 && c.colspan === 1))) {
                 a.rows.push(...b.rows.slice(1));
                 const { locations, ...first } = a.source;