@dsh-libing/knowledge-base 0.1.0 → 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/contract.d.ts

@@ -119,8 +119,12 @@ export declare const publicJobSchema: z.ZodObject<{
         modelTimeoutMs: z.ZodDefault<z.ZodNumber>;
         modelAttempts: z.ZodDefault<z.ZodNumber>;
         visionConcurrency: z.ZodDefault<z.ZodNumber>;
+        maxModelCallsLimit: z.ZodDefault<z.ZodNumber>;
+        maxModelTokensLimit: z.ZodDefault<z.ZodNumber>;
+        uploadRetentionDays: z.ZodDefault<z.ZodNumber>;
         maxModelCalls: z.ZodDefault<z.ZodNumber>;
         maxModelTokens: z.ZodDefault<z.ZodNumber>;
+        imageTimeoutMs: z.ZodDefault<z.ZodNumber>;
         visionMaxEdge: z.ZodDefault<z.ZodNumber>;
         visionTileSize: z.ZodDefault<z.ZodNumber>;
         visionMaxTiles: z.ZodDefault<z.ZodNumber>;
@@ -189,6 +193,13 @@ export declare const publicJobSchema: z.ZodObject<{
             chunkId: z.ZodString;
         }, z.core.$strip>>;
     }, z.core.$strip>;
+    pendingCalls: z.ZodDefault<z.ZodArray<z.ZodNumber>>;
+    legacyUsageCount: z.ZodDefault<z.ZodNumber>;
+    legacyCallCount: z.ZodDefault<z.ZodNumber>;
+    reportSummary: z.ZodOptional<z.ZodObject<{
+        issues: z.ZodNumber;
+        blocking: z.ZodNumber;
+    }, z.core.$strip>>;
     failures: z.ZodArray<z.ZodObject<{
         at: z.ZodString;
         stage: z.ZodEnum<{
@@ -437,8 +448,12 @@ export declare const stateSchema: z.ZodObject<{
             modelTimeoutMs: z.ZodDefault<z.ZodNumber>;
             modelAttempts: z.ZodDefault<z.ZodNumber>;
             visionConcurrency: z.ZodDefault<z.ZodNumber>;
+            maxModelCallsLimit: z.ZodDefault<z.ZodNumber>;
+            maxModelTokensLimit: z.ZodDefault<z.ZodNumber>;
+            uploadRetentionDays: z.ZodDefault<z.ZodNumber>;
             maxModelCalls: z.ZodDefault<z.ZodNumber>;
             maxModelTokens: z.ZodDefault<z.ZodNumber>;
+            imageTimeoutMs: z.ZodDefault<z.ZodNumber>;
             visionMaxEdge: z.ZodDefault<z.ZodNumber>;
             visionTileSize: z.ZodDefault<z.ZodNumber>;
             visionMaxTiles: z.ZodDefault<z.ZodNumber>;
@@ -507,6 +522,13 @@ export declare const stateSchema: z.ZodObject<{
                 chunkId: z.ZodString;
             }, z.core.$strip>>;
         }, z.core.$strip>;
+        pendingCalls: z.ZodDefault<z.ZodArray<z.ZodNumber>>;
+        legacyUsageCount: z.ZodDefault<z.ZodNumber>;
+        legacyCallCount: z.ZodDefault<z.ZodNumber>;
+        reportSummary: z.ZodOptional<z.ZodObject<{
+            issues: z.ZodNumber;
+            blocking: z.ZodNumber;
+        }, z.core.$strip>>;
         failures: z.ZodArray<z.ZodObject<{
             at: z.ZodString;
             stage: z.ZodEnum<{
@@ -557,6 +579,9 @@ export declare const stateSchema: z.ZodObject<{
         outputPath: z.ZodNullable<z.ZodString>;
         reused: z.ZodOptional<z.ZodBoolean>;
     }, z.core.$strip>>;
+    jobTotal: z.ZodDefault<z.ZodNumber>;
+    jobOffset: z.ZodDefault<z.ZodNumber>;
+    activeJobs: z.ZodDefault<z.ZodNumber>;
     dataRoot: z.ZodString;
     config: z.ZodObject<{
         chunkTargetChars: z.ZodNumber;
@@ -669,8 +694,8 @@ export declare const treeDocumentSchema: z.ZodObject<{
         reason: z.ZodOptional<z.ZodString>;
     }, z.core.$strip>>>;
     chunks: z.ZodArray<z.ZodObject<{
-        summary: z.ZodDefault<z.ZodString>;
         path: z.ZodString;
+        summary: z.ZodDefault<z.ZodString>;
         title: z.ZodString;
         headingPath: z.ZodArray<z.ZodString>;
         oversizedReason: z.ZodOptional<z.ZodString>;
@@ -972,8 +997,8 @@ export declare const treeSchema: z.ZodObject<{
             reason: z.ZodOptional<z.ZodString>;
         }, z.core.$strip>>>;
         chunks: z.ZodArray<z.ZodObject<{
-            summary: z.ZodDefault<z.ZodString>;
             path: z.ZodString;
+            summary: z.ZodDefault<z.ZodString>;
             title: z.ZodString;
             headingPath: z.ZodArray<z.ZodString>;
             oversizedReason: z.ZodOptional<z.ZodString>;
@@ -1195,3 +1220,9 @@ export declare const uploadResultSchema: z.ZodObject<{
     filename: z.ZodString;
     bytes: z.ZodNumber;
 }, z.core.$strip>;
+export declare const jobUsagePageSchema: z.ZodObject<{
+    records: z.ZodArray<z.ZodJSONSchema>;
+    offset: z.ZodNumber;
+    total: z.ZodNumber;
+    next: z.ZodNullable<z.ZodNumber>;
+}, z.core.$strip>;

package/lib/contract.js

@@ -4,7 +4,7 @@ import { jobSchema, registrySchema, librarySchema, publishedDocumentSchema, repo
 export const publicJobSchema = jobSchema.omit({ schema: true, overlays: true }).extend({ outputPath: z.string().nullable(), reused: z.boolean().optional() });
 export const locationSchema = z.object({ kbPath: z.string(), revision: z.string().nullable(), wikiPath: z.string().nullable(), indexPath: z.string().nullable() });
 export const libraryViewSchema = librarySchema.extend({ location: locationSchema, summary: z.object({ revision: z.string().nullable(), documentCount: z.number(), chunkCount: z.number(), coverageIncomplete: z.boolean() }) });
-export const stateSchema = z.object({ hostVersion: z.string(), kbs: z.array(libraryViewSchema), deletedKbs: z.array(librarySchema), activities: registrySchema.shape.activities, jobs: z.array(publicJobSchema), dataRoot: z.string(), config: z.object({ chunkTargetChars: z.number(), visionProvider: z.string(), visionModel: z.string(), maxInputBytes: z.number() }) });
+export const stateSchema = z.object({ hostVersion: z.string(), kbs: z.array(libraryViewSchema), deletedKbs: z.array(librarySchema), activities: registrySchema.shape.activities, jobs: z.array(publicJobSchema), jobTotal: z.number().int().nonnegative().default(0), jobOffset: z.number().int().nonnegative().default(0), activeJobs: z.number().int().nonnegative().default(0), dataRoot: z.string(), config: z.object({ chunkTargetChars: z.number(), visionProvider: z.string(), visionModel: z.string(), maxInputBytes: z.number() }) });
 export const treeDocumentSchema = publishedDocumentSchema.extend({ chunks: z.array(publishedDocumentSchema.shape.chunks.element.omit({ segments: true, anchors: true })), figures: z.array(publishedDocumentSchema.shape.figures.element.extend({ expectedHash: z.string().optional() })) });
 export const treeSchema = z.object({ kbId: z.string(), revision: z.string(), location: locationSchema.optional(), previewJobId: z.string().optional(), coverageIncomplete: z.boolean(), report: reportSchema, documents: z.array(treeDocumentSchema) });
 export const readSchema = z.object({ kbId: z.string(), chunkId: z.string(), revision: z.string(), text: z.string(), complete: z.boolean(), truncated: z.boolean().optional(), previewJobId: z.string().optional(),
@@ -16,3 +16,4 @@ export const modelsSchema = z.object({ providers: z.array(z.object({ id: z.strin
 export const schemaResultSchema = z.object({ content: z.string(), hash: z.string() });
 export const inspectSchema = z.object({ sourceFormat: z.literal('pdf'), pageCount: z.number(), outlineCount: z.number(), samples: z.array(z.object({ page: z.number(), chars: z.number(), hasGraphics: z.boolean(), possibleOcr: z.boolean(), hasStructure: z.boolean() })), possibleOcr: z.boolean(), note: z.string() });
 export const uploadResultSchema = z.object({ receiptId: z.string(), filename: z.string(), bytes: z.number() });
+export const jobUsagePageSchema = z.object({ records: z.array(z.json()), offset: z.number().int().nonnegative(), total: z.number().int().nonnegative(), next: z.number().int().nonnegative().nullable() });

package/lib/data.d.ts

@@ -1528,8 +1528,12 @@ export declare const jobSchema: z.ZodObject<{
         modelTimeoutMs: z.ZodDefault<z.ZodNumber>;
         modelAttempts: z.ZodDefault<z.ZodNumber>;
         visionConcurrency: z.ZodDefault<z.ZodNumber>;
+        maxModelCallsLimit: z.ZodDefault<z.ZodNumber>;
+        maxModelTokensLimit: z.ZodDefault<z.ZodNumber>;
+        uploadRetentionDays: z.ZodDefault<z.ZodNumber>;
         maxModelCalls: z.ZodDefault<z.ZodNumber>;
         maxModelTokens: z.ZodDefault<z.ZodNumber>;
+        imageTimeoutMs: z.ZodDefault<z.ZodNumber>;
         visionMaxEdge: z.ZodDefault<z.ZodNumber>;
         visionTileSize: z.ZodDefault<z.ZodNumber>;
         visionMaxTiles: z.ZodDefault<z.ZodNumber>;
@@ -1599,6 +1603,13 @@ export declare const jobSchema: z.ZodObject<{
             chunkId: z.ZodString;
         }, z.core.$strip>>;
     }, z.core.$strip>;
+    pendingCalls: z.ZodDefault<z.ZodArray<z.ZodNumber>>;
+    legacyUsageCount: z.ZodDefault<z.ZodNumber>;
+    legacyCallCount: z.ZodDefault<z.ZodNumber>;
+    reportSummary: z.ZodOptional<z.ZodObject<{
+        issues: z.ZodNumber;
+        blocking: z.ZodNumber;
+    }, z.core.$strip>>;
     usage: z.ZodObject<{
         calls: z.ZodNumber;
         tokens: z.ZodNumber;

package/lib/data.js

@@ -69,6 +69,8 @@ export const jobSchema = z.object({ jobId: idSchema, kbId: idSchema, docId: idSc
     progress: z.object({ totalPages: z.number().optional(), parsedPages: z.number().optional(), ocrPages: z.number().optional(), tables: z.number().optional(), blocks: z.number().optional(), currentPage: z.number().optional(),
         totalFigures: z.number().optional(), completedFigures: z.number().optional(), inflightFigures: z.array(z.object({ occurrenceId: z.string(), caption: z.string() })).optional(),
         totalChunks: z.number().optional(), indexedChunks: z.number().optional(), currentChunk: z.object({ title: z.string(), headingPath: z.array(z.string()), chunkId: z.string() }).optional() }),
+    pendingCalls: z.array(z.number().int().positive()).max(16).default([]), legacyUsageCount: z.number().int().nonnegative().default(0), legacyCallCount: z.number().int().nonnegative().default(0),
+    reportSummary: z.object({ issues: z.number().int().nonnegative(), blocking: z.number().int().nonnegative() }).optional(),
     usage: usageSchema, failures: z.array(z.object({ at: z.string(), stage: jobStateSchema, error: z.string() })), events: z.array(z.object({ at: z.string(), state: jobStateSchema })).optional(),
     figureRefresh: z.object({ occurrenceId: z.string(), token: z.string() }).nullable().optional(), visionCapability: z.string().optional(), parserVersion: z.string().optional(),
     pageStart: z.number().optional(), pageEnd: z.number().optional(), pdf: z.object({ pageCount: z.number(), selectedPages: rangeSchema, outlineCount: z.number().optional() }).optional(),

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');
             }