@customize-agent/knowledge 4.0.1 → 4.0.3

package/dist/extraction/content-extractor.js

@@ -5,6 +5,7 @@ import { tmpdir } from 'node:os';
 import { pathToFileURL } from 'node:url';
 import { ExternalExtractorRegistry } from './external-extractor.js';
 import { resolveAndImport, resolvePackage, getNodeModulesRoot } from './module-resolver.js';
+/** 文件内容提取器，支持文档、表格、图片、CAD 等多种文件格式的内容抽取 */
 export class ContentExtractor {
     externalExtractors;
     constructor(externalExtractors = ExternalExtractorRegistry.fromEnvironment()) {
@@ -69,12 +70,6 @@ export class ContentExtractor {
                 Object.assign(metadata, result.metadata);
                 warnings.push(...result.warnings);
             }
-            else if (file.category === 'archive') {
-                const result = await this.extractArchive(file);
-                text = result.text;
-                Object.assign(metadata, result.metadata);
-                warnings.push(...result.warnings);
-            }
             else if (file.category === 'image' && file.format !== 'vector') {
                 const result = await this.extractRasterImage(file);
                 text = result.text;
@@ -208,7 +203,7 @@ export class ContentExtractor {
         if (file.format === 'autocad' && ext === '.dxf') {
             const raw = fs.readFileSync(file.absolutePath, 'utf8');
             const layers = this.matchAll(raw, /\n\s*8\s*\n([^\n]+)/gu).slice(0, 300);
-            const textEntities = this.matchAll(raw, /\n\s*(?:1|3)\s*\n([^\n]+)/gu).slice(0, 500);
+            const textEntities = this.extractDxfTextAnnotations(raw).slice(0, 500);
             const blocks = this.matchAll(raw, /\n\s*2\s*\n([^\n]+)/gu).slice(0, 300);
             const entityTypes = this.matchAll(raw, /\n\s*0\s*\n([A-Z][A-Z0-9_]+)/gu).slice(0, 1000);
             const uniqueLayers = Array.from(new Set(layers));
@@ -221,14 +216,16 @@ export class ContentExtractor {
             metadata.blockNames = uniqueBlocks.slice(0, 80);
             metadata.entityTypeCount = uniqueEntityTypes.length;
             metadata.entityTypes = uniqueEntityTypes.slice(0, 80);
-            metadata.contentCoverage = 'dxf_layers_blocks_entities_text';
+            metadata.contentCoverage = 'dxf_semantic_layer_block_annotations';
+            const semanticNodes = this.buildCadSemanticNodes(file, uniqueLayers, uniqueBlocks, uniqueEntityTypes, textEntities);
             return {
                 text: [
                     this.metadataOnlyText(file),
                     `CAD DXF 图层: ${uniqueLayers.join(', ')}`,
                     `CAD DXF 块/符号: ${uniqueBlocks.join(', ')}`,
                     `CAD DXF 实体类型: ${uniqueEntityTypes.join(', ')}`,
-                    `CAD DXF 标注/文本:\n${textEntities.join('\n')}`,
+                    'CAD 语义图纸节点:',
+                    ...semanticNodes,
                 ].join('\n'),
                 metadata,
                 warnings,
@@ -307,7 +304,7 @@ export class ContentExtractor {
             warnings.push('dxf-parser 解析失败，已使用 DXF 文本结构抽取回退');
         }
         const layers = this.matchAll(raw, /\n\s*8\s*\n([^\n]+)/gu).slice(0, 300);
-        const textEntities = this.matchAll(raw, /\n\s*(?:1|3)\s*\n([^\n]+)/gu).slice(0, 800);
+        const textEntities = this.extractDxfTextAnnotations(raw).slice(0, 800);
         const blocks = this.matchAll(raw, /\n\s*2\s*\n([^\n]+)/gu).slice(0, 300);
         const entityTypes = this.matchAll(raw, /\n\s*0\s*\n([A-Z][A-Z0-9_]+)/gu).slice(0, 1200);
         const uniqueLayers = Array.from(new Set(layers));
@@ -320,20 +317,71 @@ export class ContentExtractor {
         metadata.blockNames = uniqueBlocks.slice(0, 80);
         metadata.entityTypeCount = uniqueEntityTypes.length;
         metadata.entityTypes = uniqueEntityTypes.slice(0, 80);
-        metadata.contentCoverage = 'dxf_layers_blocks_entities_text';
+        metadata.contentCoverage = 'dxf_semantic_layer_block_annotations';
         metadata.parsedByDxfParser = Boolean(parsed);
+        const semanticNodes = this.buildCadSemanticNodes(file, uniqueLayers, uniqueBlocks, uniqueEntityTypes, textEntities);
         return {
             text: [
                 this.metadataOnlyText(file),
                 `CAD DXF 图层: ${uniqueLayers.join(', ')}`,
                 `CAD DXF 块/符号: ${uniqueBlocks.join(', ')}`,
                 `CAD DXF 实体类型: ${uniqueEntityTypes.join(', ')}`,
-                `CAD DXF 标注/文本:\n${textEntities.join('\n')}`,
+                'CAD 语义图纸节点:',
+                ...semanticNodes,
             ].join('\n'),
             metadata,
             warnings,
         };
     }
+    buildCadSemanticNodes(file, layers, blocks, entityTypes, texts) {
+        const fileName = path.basename(file.relativePath);
+        const defaultLayer = layers[0] ?? '未命名图层';
+        const defaultBlock = blocks[0] ?? '全局模型空间';
+        const defaultEntity = entityTypes.find(type => /DIMENSION|TEXT|MTEXT|LEADER/u.test(type)) ?? entityTypes[0] ?? 'UNKNOWN';
+        const annotations = texts.length > 0 ? texts : [{ text: '未提取到文字标注', layer: defaultLayer, block: defaultBlock, entityType: defaultEntity }];
+        return annotations.map((annotation, index) => {
+            const nearest = this.findNearestCadAnnotation(annotation, annotations);
+            const layer = annotation.layer ?? layers[index % Math.max(1, layers.length)] ?? defaultLayer;
+            const block = annotation.block ?? blocks[index % Math.max(1, blocks.length)] ?? defaultBlock;
+            const entity = annotation.entityType ?? this.inferCadEntityType(annotation.text, defaultEntity);
+            const status = /关键|critical|尺寸|dim|mm|cm|m\b|°|φ|Φ|R\d/iu.test(annotation.text) ? '关键尺寸/约束候选' : '普通标注';
+            const position = annotation.x != null && annotation.y != null ? ` | 坐标: (${annotation.x.toFixed(2)}, ${annotation.y.toFixed(2)})` : '';
+            return [`图纸节点: ${fileName} | 图层: ${layer} | 块: ${block} | 实体类型: ${entity}${position}`, `└── 标注文本: ${annotation.text} | 关联对象: ${nearest ? `邻近标注 ${nearest.text}` : '空间邻近候选'} | 状态: ${status}`].join('\n');
+        });
+    }
+    extractDxfTextAnnotations(raw) {
+        const entities = raw.split(/\n\s*0\s*\n/u).filter(section => /^(?:TEXT|MTEXT|DIMENSION|LEADER)/u.test(section.trim()));
+        return entities.flatMap(section => {
+            const text = /\n\s*(?:1|3)\s*\n([^\n]+)/u.exec(section)?.[1]?.trim();
+            if (!text)
+                return [];
+            return [{
+                    text,
+                    layer: /\n\s*8\s*\n([^\n]+)/u.exec(section)?.[1]?.trim(),
+                    block: /\n\s*2\s*\n([^\n]+)/u.exec(section)?.[1]?.trim(),
+                    entityType: section.trim().split(/\s+/u)[0],
+                    x: Number(/\n\s*10\s*\n([^\n]+)/u.exec(section)?.[1]),
+                    y: Number(/\n\s*20\s*\n([^\n]+)/u.exec(section)?.[1]),
+                }].map(item => ({ ...item, x: Number.isFinite(item.x) ? item.x : undefined, y: Number.isFinite(item.y) ? item.y : undefined }));
+        });
+    }
+    findNearestCadAnnotation(target, annotations) {
+        if (target.x == null || target.y == null)
+            return undefined;
+        return annotations
+            .filter(item => item !== target && item.x != null && item.y != null)
+            .map(item => ({ item, distance: Math.hypot((item.x ?? 0) - target.x, (item.y ?? 0) - target.y) }))
+            .sort((a, b) => a.distance - b.distance)[0]?.item;
+    }
+    inferCadEntityType(text, fallback) {
+        if (/\b\d+(?:\.\d+)?\s*(?:mm|cm|m)\b|φ|Φ|R\d/iu.test(text))
+            return '线性尺寸/半径尺寸';
+        if (/°|angle|角度/iu.test(text))
+            return '角度尺寸';
+        if (/note|说明|备注/iu.test(text))
+            return '文字说明';
+        return fallback;
+    }
     async tryConvertDwgWithBundledWasm(filePath) {
         try {
             const mod = await resolveAndImport('dwgdxf');
@@ -447,36 +495,39 @@ export class ContentExtractor {
     }
     extractData(file) {
         const raw = fs.readFileSync(file.absolutePath, 'utf8');
-        const metadata = { extractionMode: 'structured_data', vectorizable: true };
+        const metadata = { extractionMode: 'structured_data', semanticExtractionMode: 'structured_data_semantic_paths', vectorizable: true };
         try {
             if (file.format === 'json') {
                 const isJsonl = path.extname(file.absolutePath).toLowerCase() === '.jsonl';
-                const lines = isJsonl
-                    ? raw.split(/\r?\n/u).filter(Boolean).slice(0, 500).flatMap((line, index) => this.flattenJson(JSON.parse(line), `line${index + 1}`))
-                    : this.flattenJson(JSON.parse(raw));
+                const records = isJsonl
+                    ? raw.split(/\r?\n/u).filter(Boolean).map((line, index) => ({ path: `line${index + 1}`, value: JSON.parse(line) }))
+                    : [{ path: '$', value: JSON.parse(raw) }];
+                const lines = records.flatMap(record => this.flattenJson(record.value, record.path));
+                const objects = records.flatMap(record => this.atomicJsonObjects(record.value, record.path));
                 metadata.fieldCount = lines.length;
-                metadata.recordCount = isJsonl ? raw.split(/\r?\n/u).filter(Boolean).length : 1;
+                metadata.recordCount = records.length;
+                metadata.objectCount = objects.length;
                 metadata.dataPaths = lines.map(line => line.split(':')[0]).slice(0, 200);
-                metadata.contentCoverage = isJsonl ? 'jsonl_records_paths_values' : 'json_paths_values';
-                return { text: [this.metadataOnlyText(file), ...lines.slice(0, 1500)].join('\n'), metadata, warnings: [] };
+                metadata.contentCoverage = isJsonl ? 'jsonl_atomic_objects_paths_values' : 'json_atomic_objects_paths_values';
+                return { text: [this.metadataOnlyText(file), '## 路径声明', ...lines, '## 原子对象', ...objects].join('\n'), metadata, warnings: [] };
             }
         }
         catch {
             metadata.parseError = true;
         }
         if (file.format === 'yaml') {
-            const pairs = this.matchAll(raw, /^\s*([\w.-]+)\s*:\s*(.{1,300})$/gmu).slice(0, 1500);
-            metadata.fieldCount = pairs.length;
-            metadata.dataPaths = pairs.map(line => (line.split(':')[0] ?? '').trim()).slice(0, 200);
-            metadata.contentCoverage = 'yaml_key_values';
-            return { text: [this.metadataOnlyText(file), ...pairs].join('\n'), metadata, warnings: [] };
+            const lines = this.flattenYamlByIndent(raw).slice(0, 3000);
+            metadata.fieldCount = lines.length;
+            metadata.dataPaths = lines.map(line => (line.split(':')[0] ?? '').trim()).slice(0, 200);
+            metadata.contentCoverage = 'yaml_indented_paths_values';
+            return { text: [this.metadataOnlyText(file), '## YAML 路径声明', ...lines].join('\n'), metadata, warnings: [] };
         }
         if (file.format === 'xml') {
-            const elements = this.matchAll(raw, /<([A-Za-z_][\w:.-]*)\b[^>]*>([^<]{1,200})<\/\1>/gu).slice(0, 1000);
+            const elements = this.flattenXmlPaths(raw).slice(0, 3000);
             metadata.elementTextCount = elements.length;
-            metadata.dataPaths = elements.map(line => line.match(/^<([A-Za-z_][\w:.-]*)/u)?.[1]).filter(Boolean).slice(0, 200);
-            metadata.contentCoverage = 'xml_element_text';
-            return { text: [this.metadataOnlyText(file), ...elements].join('\n'), metadata, warnings: [] };
+            metadata.dataPaths = elements.map(line => (line.split(':')[0] ?? '').trim()).slice(0, 200);
+            metadata.contentCoverage = 'xml_paths_values';
+            return { text: [this.metadataOnlyText(file), '## XML 路径声明', ...elements].join('\n'), metadata, warnings: [] };
         }
         metadata.contentCoverage = 'plain_structured_text';
         return { text: [this.metadataOnlyText(file), raw].join('\n'), metadata, warnings: [] };
@@ -485,18 +536,19 @@ export class ContentExtractor {
         const raw = fs.readFileSync(file.absolutePath, 'utf8');
         const metadata = { extractionMode: 'diagram_structural', vectorizable: true };
         if (file.format === 'drawio') {
-            const labels = this.matchAll(raw, /(?:value|label)="([^"]+)"/gu).map(value => this.stripXml(value)).slice(0, 300);
-            metadata.nodeTextCount = labels.length;
-            metadata.contentCoverage = 'drawio_labels';
-            return { text: [this.metadataOnlyText(file), `Draw.io 节点/连线文本:\n${labels.join('\n')}`].join('\n'), metadata, warnings: [] };
+            const graph = this.extractDrawioGraph(raw);
+            metadata.nodeTextCount = graph.nodes.length;
+            metadata.edgeCount = graph.edges.length;
+            metadata.contentCoverage = 'drawio_graph_links';
+            return { text: [this.metadataOnlyText(file), 'Draw.io 图链路:', ...graph.links, 'Draw.io 节点:', ...graph.nodes.map(node => `[${node.id}] ${node.label}`)].join('\n'), metadata, warnings: [] };
         }
         if (file.format === 'excalidraw') {
             try {
-                const parsed = JSON.parse(raw);
-                const texts = (parsed.elements ?? []).filter(element => element.text).map(element => `${element.type ?? 'shape'}: ${element.text}`).slice(0, 300);
-                metadata.elementTextCount = texts.length;
-                metadata.contentCoverage = 'excalidraw_text_elements';
-                return { text: [this.metadataOnlyText(file), `Excalidraw 图形文本:\n${texts.join('\n')}`].join('\n'), metadata, warnings: [] };
+                const graph = this.extractExcalidrawGraph(raw);
+                metadata.elementTextCount = graph.nodes.length;
+                metadata.edgeCount = graph.links.length;
+                metadata.contentCoverage = 'excalidraw_graph_links';
+                return { text: [this.metadataOnlyText(file), 'Excalidraw 图链路:', ...graph.links, 'Excalidraw 节点:', ...graph.nodes].join('\n'), metadata, warnings: [] };
             }
             catch {
                 metadata.parseError = true;
@@ -505,6 +557,52 @@ export class ContentExtractor {
         metadata.contentCoverage = 'diagram_source_text';
         return { text: [this.metadataOnlyText(file), raw].join('\n'), metadata, warnings: [] };
     }
+    extractDrawioGraph(raw) {
+        const cells = Array.from(raw.matchAll(/<mxCell\b([^>]*)>/gu), match => this.parseXmlAttributes(match[1] ?? ''));
+        const nodes = cells
+            .filter(cell => cell.value || cell.label || cell.id)
+            .map((cell, index) => ({ id: String(cell.id ?? `node-${index + 1}`), label: this.stripXml(String(cell.value || cell.label || cell.id)) }))
+            .filter(node => node.label);
+        const nodeById = new Map(nodes.map(node => [node.id, node.label]));
+        const edges = cells.filter(cell => cell.edge === '1' && cell.source && cell.target);
+        const links = edges.map(edge => {
+            const source = nodeById.get(String(edge.source)) ?? String(edge.source);
+            const target = nodeById.get(String(edge.target)) ?? String(edge.target);
+            const label = this.stripXml(String(edge.value || edge.label || '关系'));
+            return `[${source}] ──> (${label}) ──> [${target}]`;
+        });
+        return { nodes, edges: edges.map(edge => String(edge.id ?? 'edge')), links };
+    }
+    extractExcalidrawGraph(raw) {
+        const parsed = JSON.parse(raw);
+        const elements = parsed.elements ?? [];
+        const textById = new Map(elements.filter(element => element.text).map(element => [String(element.id), String(element.text)]));
+        const shapeLabels = elements
+            .filter(element => element.type !== 'arrow' && element.type !== 'line')
+            .map((element, index) => {
+            const id = String(element.id ?? `node-${index + 1}`);
+            const own = element.text ?? textById.get(id);
+            const nested = elements.find(candidate => candidate.text && this.isPointInside(candidate, element));
+            return { id, label: own ?? nested?.text ?? element.type ?? '节点' };
+        });
+        const labelById = new Map(shapeLabels.map(node => [node.id, node.label]));
+        const links = elements
+            .filter(element => element.type === 'arrow' && element.startBinding?.elementId && element.endBinding?.elementId)
+            .map(element => `[${labelById.get(element.startBinding.elementId) ?? element.startBinding.elementId}] ──> (箭头) ──> [${labelById.get(element.endBinding.elementId) ?? element.endBinding.elementId}]`);
+        const nodes = shapeLabels.map(node => `[${node.id}] ${node.label}`);
+        return { nodes, links };
+    }
+    isPointInside(point, box) {
+        if (point.x == null || point.y == null || box.x == null || box.y == null || box.width == null || box.height == null)
+            return false;
+        return point.x >= box.x && point.x <= box.x + box.width && point.y >= box.y && point.y <= box.y + box.height;
+    }
+    parseXmlAttributes(input) {
+        const attrs = {};
+        for (const match of input.matchAll(/([\w:-]+)="([^"]*)"/gu))
+            attrs[match[1]] = this.stripXml(match[2] ?? '');
+        return attrs;
+    }
     parseDelimitedLine(line, delimiter) {
         const values = [];
         let current = '';
@@ -531,28 +629,38 @@ export class ContentExtractor {
         values.push(current.trim());
         return values;
     }
+    toMarkdownTable(header, rows) {
+        const width = Math.max(header.length, ...rows.map(row => row.length), 1);
+        const normalizedHeader = Array.from({ length: width }, (_, index) => header[index] || `COL${index + 1}`);
+        const escape = (value) => String(value ?? '').replace(/\|/gu, '\\|').replace(/\r?\n/gu, ' ').trim();
+        return [
+            `| ${normalizedHeader.map(escape).join(' | ')} |`,
+            `| ${normalizedHeader.map(() => '---').join(' | ')} |`,
+            ...rows.map(row => `| ${Array.from({ length: width }, (_, index) => escape(row[index] ?? '')).join(' | ')} |`),
+        ].join('\n');
+    }
     extractDelimitedText(file) {
         const raw = fs.readFileSync(file.absolutePath, 'utf8');
         const delimiter = file.format === 'tsv' ? '\t' : ',';
         const rows = raw.split(/\r?\n/u).filter(line => line.trim().length > 0);
         const header = rows[0] ? this.parseDelimitedLine(rows[0], delimiter) : [];
-        const structured = rows.slice(1, 501).flatMap((line, rowIndex) => {
-            const values = this.parseDelimitedLine(line, delimiter);
-            return values.map((value, colIndex) => {
-                const column = header[colIndex] || `COL${colIndex + 1}`;
-                return `R${rowIndex + 2}C${colIndex + 1} ${column}: ${value}`;
-            });
-        });
+        const tableRows = rows.slice(1).map(line => this.parseDelimitedLine(line, delimiter));
+        const markdown = this.toMarkdownTable(header, tableRows);
+        const legacyKv = tableRows.flatMap((values, rowIndex) => values.map((value, colIndex) => {
+            const column = header[colIndex] || `COL${colIndex + 1}`;
+            return `R${rowIndex + 2}C${colIndex + 1} ${column}: ${value}`;
+        }));
         return {
-            text: [this.metadataOnlyText(file), `表头: ${header.join(' | ')}`, ...structured, raw].join('\n'),
+            text: [this.metadataOnlyText(file), '### 工作表: 默认', markdown, '### 表格路径声明', ...legacyKv].join('\n\n'),
             metadata: {
                 extractionMode: 'delimited_text_structured',
+                semanticExtractionMode: 'delimited_markdown_table',
                 vectorizable: true,
                 delimiter: file.format === 'tsv' ? 'tab' : 'comma',
                 rowCount: rows.length,
                 columnCount: header.length,
                 columnNames: header.slice(0, 120),
-                contentCoverage: 'table_headers_cells_text',
+                contentCoverage: 'markdown_table',
             },
             warnings: [],
         };
@@ -567,19 +675,27 @@ export class ContentExtractor {
             return this.extractLegacyOfficeBinary(file);
         if (ext === '.docx') {
             try {
+                const styledMarkdown = await this.extractDocxStyleTreeMarkdown(file.absolutePath);
+                if (styledMarkdown.trim()) {
+                    return {
+                        text: styledMarkdown,
+                        metadata: { extractionMode: 'docx_xml_style_tree_markdown', vectorizable: true, contentCoverage: 'office_style_tree_markdown' },
+                        warnings: [],
+                    };
+                }
                 const mammoth = await resolveAndImport('mammoth');
-                const result = await mammoth.extractRawText({ path: file.absolutePath });
-                const text = result.value.trim();
+                const result = await mammoth.convertToMarkdown({ path: file.absolutePath });
+                const text = String(result.value ?? '').trim();
                 if (text) {
                     return {
-                        text,
-                        metadata: { extractionMode: 'builtin_mammoth', vectorizable: true, contentCoverage: 'office_full_text' },
+                        text: this.normalizeMarkdownHeadings(text),
+                        metadata: { extractionMode: 'builtin_mammoth_markdown', vectorizable: true, contentCoverage: 'office_markdown_structure' },
                         warnings: result.messages.map(m => m.message),
                     };
                 }
             }
             catch {
-                // fallback below
+                // 降级到下方 Office ZIP 解析
             }
         }
         return this.extractOfficeZip(file);
@@ -630,6 +746,14 @@ export class ContentExtractor {
             warnings: text ? [] : ['旧版 Office 二进制文件未提取到正文，未入库'],
         };
     }
+    findMergedCellValue(sheet, merges, row, col, XLSX) {
+        const merge = merges.find(item => row >= item.s.r && row <= item.e.r && col >= item.s.c && col <= item.e.c);
+        if (!merge)
+            return undefined;
+        const originAddress = XLSX.utils.encode_cell({ r: merge.s.r, c: merge.s.c });
+        const origin = sheet[originAddress];
+        return origin?.w ?? (origin?.v == null ? undefined : String(origin.v));
+    }
     async extractSpreadsheet(file) {
         const ext = path.extname(file.absolutePath).toLowerCase();
         try {
@@ -644,28 +768,32 @@ export class ContentExtractor {
                 if (!sheet || !sheet['!ref'])
                     continue;
                 const range = XLSX.utils.decode_range(sheet['!ref']);
-                const lines = [`工作表: ${name}`, `范围: ${sheet['!ref']}`];
                 const merges = sheet['!merges'] ?? [];
                 mergeCount += merges.length;
-                if (merges.length > 0) {
-                    lines.push(`合并单元格: ${merges.map(item => `${XLSX.utils.encode_cell(item.s)}:${XLSX.utils.encode_cell(item.e)}`).join(', ')}`);
-                }
+                const matrix = [];
                 for (let row = range.s.r; row <= range.e.r; row++) {
+                    const values = [];
                     for (let col = range.s.c; col <= range.e.c; col++) {
                         const address = XLSX.utils.encode_cell({ r: row, c: col });
                         const cell = sheet[address];
-                        if (!cell || (cell.v == null && !cell.f))
-                            continue;
-                        cellCount++;
-                        if (cell.f)
-                            formulaCount++;
-                        const display = cell.w ?? String(cell.v ?? '');
-                        const formula = cell.f ? ` 公式=${cell.f}` : '';
-                        lines.push(`${address}: ${display}${formula}`);
+                        const mergedValue = cell ? undefined : this.findMergedCellValue(sheet, merges, row, col, XLSX);
+                        if (cell) {
+                            cellCount++;
+                            if (cell.f)
+                                formulaCount++;
+                        }
+                        const display = cell?.w ?? String(cell?.v ?? mergedValue ?? '');
+                        const formula = cell?.f ? ` 公式=${cell.f}` : '';
+                        values.push(`${display}${formula}`.trim());
                     }
+                    if (values.some(Boolean))
+                        matrix.push(values);
+                }
+                if (matrix.length > 0) {
+                    const header = matrix[0] ?? [];
+                    const rows = matrix.slice(1);
+                    sheetTexts.push([`### 工作表: ${name}`, this.toMarkdownTable(header, rows)].join('\n\n'));
                 }
-                if (lines.length > 2)
-                    sheetTexts.push(lines.slice(0, 5_000).join('\n'));
             }
             if (sheetTexts.length > 0) {
                 return {
@@ -678,12 +806,47 @@ export class ContentExtractor {
         catch {
             if (ext === '.xls')
                 return this.extractLegacyOfficeBinary(file);
-            // fallback below
+            // 降级到下方 ZIP 文件内容提取
         }
         if (ext === '.xls')
             return this.extractLegacyOfficeBinary(file);
         return this.extractOfficeZip(file);
     }
+    async extractDocxStyleTreeMarkdown(filePath) {
+        const jszipMod = await resolveAndImport('jszip');
+        const JSZip = jszipMod.default ?? jszipMod;
+        const zip = await JSZip.loadAsync(fs.readFileSync(filePath));
+        const docXml = await zip.files['word/document.xml']?.async('string');
+        if (!docXml)
+            return '';
+        const paragraphs = Array.from(docXml.matchAll(/<w:p[\s\S]*?<\/w:p>/gu), match => match[0]);
+        const lines = [];
+        for (const paragraph of paragraphs) {
+            const texts = Array.from(paragraph.matchAll(/<w:t[^>]*>([\s\S]*?)<\/w:t>/gu), match => this.stripXml(match[1] ?? '')).join('');
+            if (!texts.trim())
+                continue;
+            const style = /<w:pStyle\s+w:val="([^"]+)"/u.exec(paragraph)?.[1] ?? '';
+            const bold = /<w:b\b/u.test(paragraph);
+            const size = Number(/<w:sz\s+w:val="(\d+)"/u.exec(paragraph)?.[1] ?? 0);
+            const level = this.docxHeadingLevel(style, bold, size, texts);
+            lines.push(`${level > 0 ? `${'#'.repeat(level)} ` : ''}${texts.trim()}`);
+        }
+        return lines.join('\n\n');
+    }
+    docxHeadingLevel(style, bold, size, text) {
+        const normalized = style.toLowerCase();
+        const heading = /heading(\d)|标题(\d)|h(\d)/iu.exec(normalized);
+        const styleLevel = Number(heading?.[1] ?? heading?.[2] ?? heading?.[3] ?? 0);
+        if (styleLevel >= 1 && styleLevel <= 6)
+            return styleLevel;
+        if (text.length <= 100 && size >= 32)
+            return 1;
+        if (text.length <= 100 && size >= 28)
+            return 2;
+        if (text.length <= 100 && (size >= 24 || bold))
+            return 3;
+        return 0;
+    }
     async extractOfficeZip(file) {
         const metadata = { extractionMode: 'office_zip_text', vectorizable: true };
         try {
@@ -709,26 +872,6 @@ export class ContentExtractor {
             return { text: '', metadata, warnings: [`Office/表格/演示文件解析失败: ${metadata.parseError}，未入库`] };
         }
     }
-    async extractArchive(file) {
-        const metadata = { extractionMode: 'archive_manifest', vectorizable: true };
-        if (path.extname(file.absolutePath).toLowerCase() !== '.zip') {
-            metadata.contentCoverage = 'metadata_filename';
-            return { text: this.metadataOnlyText(file), metadata, warnings: ['压缩包未提取到正文，未入库；仅 zip 可提取文件清单'] };
-        }
-        try {
-            const jszipMod = await resolveAndImport('jszip');
-            const JSZip = jszipMod.default ?? jszipMod;
-            const zip = await JSZip.loadAsync(fs.readFileSync(file.absolutePath));
-            const entries = Object.values(zip.files).map(entry => `${entry.dir ? '目录' : '文件'}: ${entry.name}`).slice(0, 1_000);
-            metadata.entryCount = Object.keys(zip.files).length;
-            metadata.contentCoverage = 'zip_manifest';
-            return { text: [this.metadataOnlyText(file), '压缩包文件清单:', ...entries].join('\n'), metadata, warnings: [] };
-        }
-        catch (error) {
-            metadata.parseError = error instanceof Error ? error.message : String(error);
-            return { text: '', metadata, warnings: ['压缩包解析失败，内置解析器未提取到文件清单，未入库'] };
-        }
-    }
     async extractRasterImage(file) {
         const metadata = { extractionMode: 'builtin_tesseract_ocr_isolated', vectorizable: true };
         if (process.env.CUSTOMIZE_AGENT_DISABLE_OCR === '1') {
@@ -742,6 +885,13 @@ export class ContentExtractor {
             metadata.parseError = validationError;
             return { text: '', metadata, warnings: [`图片文件无效或不完整：${validationError}，未入库`] };
         }
+        const paddle = await this.tryPaddleOcrLayout(file.absolutePath);
+        if (paddle) {
+            metadata.contentCoverage = 'paddleocr_layout_regions';
+            metadata.ocrProvider = 'paddleocr';
+            metadata.ocrRegionCount = paddle.regionCount;
+            return { text: [this.metadataOnlyText(file), paddle.text].join('\n'), metadata, warnings: [] };
+        }
         let tesseractPath;
         try {
             tesseractPath = resolvePackage('tesseract.js');
@@ -766,7 +916,12 @@ const { createWorker } = tesseractMod;
 const worker = await createWorker('chi_sim+eng');
 try {
   const result = await worker.recognize(process.argv[1]);
-  process.stdout.write(result.data.text || '');
+  const lines = (result.data.lines || []).map((line, index) => ({
+    index: index + 1,
+    text: line.text || '',
+    bbox: line.bbox || line.baseline || null,
+  })).filter(line => line.text.trim());
+  process.stdout.write(JSON.stringify({ text: result.data.text || '', lines }));
 } finally {
   await worker.terminate();
 }`,
@@ -778,17 +933,73 @@ try {
             metadata.parseError = message;
             return { text: '', metadata, warnings: [`内置 OCR 解析失败：${message}，未入库`] };
         }
-        const text = result.stdout.trim();
-        metadata.contentCoverage = text ? 'ocr_text' : 'metadata_filename';
+        const parsed = this.parseOcrJson(result.stdout);
+        const text = parsed.text.trim();
+        const regions = this.formatOcrRegions(parsed.lines);
+        metadata.contentCoverage = text ? 'ocr_text_bounding_boxes' : 'metadata_filename';
         metadata.ocrProvider = 'tesseract.js';
         metadata.ocrLanguages = 'chi_sim+eng';
         metadata.ocrTextLength = text.length;
+        metadata.ocrLineCount = parsed.lines.length;
         return {
-            text: text ? [this.metadataOnlyText(file), `OCR 识别文本:\n${text}`].join('\n') : '',
+            text: text ? [this.metadataOnlyText(file), 'OCR 区域文本:', ...regions, `OCR 完整文本:\n${text}`].join('\n') : '',
             metadata,
             warnings: text ? [] : ['内置 OCR 未识别到文字，未入库'],
         };
     }
+    async tryPaddleOcrLayout(filePath) {
+        const command = process.env.CUSTOMIZE_PADDLE_OCR_CMD || process.env.PADDLE_OCR_CMD;
+        if (!command)
+            return undefined;
+        const result = spawnSync(command, [filePath], { encoding: 'utf8', timeout: 0, maxBuffer: 50 * 1024 * 1024, shell: true });
+        if (result.status !== 0 || !result.stdout.trim())
+            return undefined;
+        try {
+            const parsed = JSON.parse(result.stdout);
+            const lines = parsed.map((region, index) => `区域 ${index + 1} [${region.type ?? 'text'}] ${this.formatBoundingBox(region.bbox)}: ${region.text ?? ''}`);
+            return { text: ['OCR 版面分析区域:', ...lines].join('\n'), regionCount: lines.length };
+        }
+        catch {
+            const lines = result.stdout.split(/\r?\n/u).filter(Boolean);
+            return { text: ['OCR 版面分析区域:', ...lines].join('\n'), regionCount: lines.length };
+        }
+    }
+    parseOcrJson(raw) {
+        try {
+            const parsed = JSON.parse(raw);
+            return {
+                text: parsed.text ?? raw,
+                lines: (parsed.lines ?? []).map((line, index) => ({ index: line.index ?? index + 1, text: line.text ?? '', bbox: line.bbox })).filter(line => line.text.trim()),
+            };
+        }
+        catch {
+            return { text: raw, lines: raw.split(/\r?\n/u).map((text, index) => ({ index: index + 1, text })).filter(line => line.text.trim()) };
+        }
+    }
+    formatOcrRegions(lines) {
+        return lines.map(line => {
+            const bbox = this.formatBoundingBox(line.bbox);
+            const type = this.classifyOcrRegion(line.text);
+            return `区域 ${line.index} [${type}]${bbox ? ` ${bbox}` : ''}: ${line.text}`;
+        });
+    }
+    classifyOcrRegion(text) {
+        if (/\|/.test(text) || /\s{2,}/u.test(text) || /表\s*\d|合计|小计/u.test(text))
+            return 'table';
+        if (/图\s*\d|figure|image|示意图/iu.test(text))
+            return 'image-caption';
+        return 'text';
+    }
+    formatBoundingBox(value) {
+        if (!value || typeof value !== 'object')
+            return '';
+        const record = value;
+        const x0 = record.x0 ?? record.left ?? record.x;
+        const y0 = record.y0 ?? record.top ?? record.y;
+        const x1 = record.x1 ?? record.right;
+        const y1 = record.y1 ?? record.bottom;
+        return [x0, y0, x1, y1].some(item => item != null) ? `[bbox x0=${x0 ?? ''}, y0=${y0 ?? ''}, x1=${x1 ?? ''}, y1=${y1 ?? ''}]` : '';
+    }
     validateRasterImage(filePath) {
         const buffer = fs.readFileSync(filePath);
         if (buffer.length < 12)
@@ -816,25 +1027,25 @@ try {
     async extractPdf(file) {
         const metadata = { extractionMode: 'pdf_text', vectorizable: true };
         const warnings = [];
-        // Tier 1: pdfjs-dist 文本提取（处理常规 PDF、压缩内容流、CJK 字体等）
+        // 第一层：pdfjs-dist 文本提取（处理常规 PDF、压缩内容流、CJK 字体等）
         try {
             const raw = fs.readFileSync(file.absolutePath);
             const text = await this.extractPdfText(raw);
             if (text.trim()) {
-                metadata.contentCoverage = 'pdf_text_streams';
+                metadata.contentCoverage = 'pdf_text_streams_layout_markdown';
                 metadata.pdfExtractor = 'pdfjs-dist';
-                return { text: [this.metadataOnlyText(file), text].join('\n'), metadata, warnings };
+                return { text: [this.metadataOnlyText(file), this.toMarkdownDocument(text)].join('\n\n'), metadata, warnings };
             }
         }
         catch (error) {
             warnings.push(`PDF 文本提取失败: ${error instanceof Error ? error.message : String(error)}`);
             metadata.parseError = error instanceof Error ? error.message : String(error);
         }
-        // Tier 2: OCR（扫描件/图片型 PDF）—— 必须保留并确保可用
+        // 第二层：OCR（扫描件/图片型 PDF）—— 必须保留并确保可用
         const ocr = await this.extractScannedPdfOcr(file);
         if (ocr.text.trim())
             return ocr;
-        // Tier 3: 仅索引元数据（兜底）
+        // 第三层：仅索引元数据（兜底）
         metadata.extractionMode = 'pdf_metadata_only';
         metadata.contentCoverage = 'metadata_filename';
         metadata.ocrRecommended = true;
@@ -846,10 +1057,6 @@ try {
             warnings: [...warnings, 'PDF 正文暂未提取到文本，已索引文件名、路径和类型元数据', ...ocr.warnings],
         };
     }
-    pdfOcrPageLimit() {
-        const value = Number(process.env.KB_PDF_OCR_PAGE_LIMIT ?? 20);
-        return Number.isFinite(value) ? Math.min(500, Math.max(1, Math.floor(value))) : 20;
-    }
     async extractScannedPdfOcr(file) {
         const metadata = {
             extractionMode: 'pdf_page_ocr_embedded',
@@ -858,9 +1065,8 @@ try {
             ocrProvider: 'tesseract.js',
             ocrLanguages: 'chi_sim+eng',
             pdfRenderer: 'pdfjs-dist + @napi-rs/canvas',
-            pdfOcrPageLimit: this.pdfOcrPageLimit(),
+            pdfOcrPageLimit: 'all',
         };
-        const pageLimitConfig = this.pdfOcrPageLimit();
         // 解析模块路径 —— 确保在打包 Server 等上下文中子进程也能正确加载
         let canvasPath;
         let pdfjsPath;
@@ -890,10 +1096,9 @@ import { createCanvas } from ${JSON.stringify(canvasPath)};
 import * as pdfjs from ${JSON.stringify(pdfjsPath)};
 import { createWorker } from ${JSON.stringify(tesseractPath)};
 const filePath = process.argv[1];
-const configuredLimit = Math.max(1, Number(process.argv[2] || 5));
 const bytes = new Uint8Array(fs.readFileSync(filePath));
 const doc = await pdfjs.getDocument({ data: bytes, verbosity: 0 }).promise;
-const pageLimit = Math.min(doc.numPages, configuredLimit);
+const pageLimit = doc.numPages;
 const tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'kb-pdf-ocr-'));
 const worker = await createWorker('chi_sim+eng');
 const pages = [];
@@ -908,7 +1113,8 @@ try {
     fs.writeFileSync(imagePath, canvas.toBuffer('image/png'));
     const recognized = await worker.recognize(imagePath);
     const text = (recognized.data.text || '').trim();
-    if (text) pages.push('PDF OCR 第 ' + i + ' 页:\\n' + text);
+    const lines = (recognized.data.lines || []).map((line, index) => ({ index: index + 1, text: line.text || '', bbox: line.bbox || null })).filter(line => line.text.trim());
+    if (text) pages.push('PDF OCR 第 ' + i + ' 页:\\n' + lines.map(line => '区域 ' + line.index + ' ' + JSON.stringify(line.bbox || {}) + ': ' + line.text).join('\\n') + '\\n\\n完整文本:\\n' + text);
   }
 } finally {
   await worker.terminate();
@@ -916,8 +1122,7 @@ try {
 }
 process.stdout.write(JSON.stringify({ pageCount: doc.numPages, pageLimit, text: pages.join('\\n\\n') }));`,
             file.absolutePath,
-            String(pageLimitConfig),
-        ], { encoding: 'utf8', timeout: Math.max(120_000, pageLimitConfig * 60_000), maxBuffer: 50 * 1024 * 1024, env: childEnv });
+        ], { encoding: 'utf8', timeout: 0, maxBuffer: 50 * 1024 * 1024, env: childEnv });
         if (result.status !== 0 || result.error) {
             metadata.ocrRecommended = true;
             metadata.ocrReason = result.error?.message ?? result.stderr.trim() ?? `pdf_page_ocr_exit_${result.status ?? 'unknown'}`;
@@ -927,7 +1132,7 @@ process.stdout.write(JSON.stringify({ pageCount: doc.numPages, pageLimit, text:
             const parsed = JSON.parse(result.stdout);
             const text = parsed.text?.trim() ?? '';
             metadata.ocrPageCount = parsed.pageCount ?? 0;
-            metadata.pdfOcrPageLimit = parsed.pageLimit ?? 5;
+            metadata.pdfOcrPageLimit = parsed.pageLimit ?? 'all';
             metadata.ocrTextLength = text.length;
             metadata.contentCoverage = text ? 'pdf_page_ocr_text' : 'metadata_filename';
             return { text: text ? [this.metadataOnlyText(file), text].join('\n') : '', metadata, warnings: text ? [] : ['内置扫描 PDF OCR 未识别到文字'] };
@@ -939,23 +1144,20 @@ process.stdout.write(JSON.stringify({ pageCount: doc.numPages, pageLimit, text:
         }
     }
     async extractPdfText(buffer) {
-        // Tier 1: pdfjs-dist 文本提取（处理压缩内容流、CJK 字体、现代 PDF）
+        // 第一层：pdfjs-dist 文本提取（处理压缩内容流、CJK 字体、现代 PDF）
         try {
             const mod = await resolveAndImport('pdfjs-dist/legacy/build/pdf.mjs');
             const loadingTask = mod.getDocument({ data: new Uint8Array(buffer), verbosity: 0 });
             const doc = await loadingTask.promise;
             const pages = [];
-            const pageLimit = Math.min(doc.numPages, 200);
+            const pageLimit = doc.numPages;
             for (let i = 1; i <= pageLimit; i++) {
                 const page = await doc.getPage(i);
                 const content = await page.getTextContent();
-                const pageText = content.items
-                    .map((item) => {
-                    const hasStr = item != null && typeof item === 'object' && 'str' in item;
-                    return hasStr ? String(item.str) : '';
-                })
-                    .filter((s) => s.trim().length > 0)
-                    .join(' ');
+                const items = content.items
+                    .map((item) => this.toPdfTextItem(item))
+                    .filter((item) => !!item && item.str.trim().length > 0);
+                const pageText = this.layoutPdfTextItems(items, i);
                 if (pageText.trim())
                     pages.push(pageText.trim());
             }
@@ -970,7 +1172,7 @@ process.stdout.write(JSON.stringify({ pageCount: doc.numPages, pageLimit, text:
             if (process.env.KB_DEBUG === '1')
                 console.warn('[kb] pdfjs-dist extraction failed:', e.message);
         }
-        // Tier 2: pdf-parse（兼容旧版 PDF）
+        // 第二层：pdf-parse（兼容旧版 PDF）
         try {
             const mod = await resolveAndImport('pdf-parse');
             const pdfParse = mod.default;
@@ -981,9 +1183,9 @@ process.stdout.write(JSON.stringify({ pageCount: doc.numPages, pageLimit, text:
             }
         }
         catch {
-            // fallback to raw regex below
+            // 降级到下方纯正则提取
         }
-        // Tier 3: raw regex 回退（未压缩的古老 PDF）
+        // 第三层：raw regex 回退（未压缩的古老 PDF）
         const raw = buffer.toString('latin1');
         const matches = Array.from(raw.matchAll(/\(([^()]{2,500})\)\s*T[jJ]/gu), match => match[1] ?? '')
             .concat(Array.from(raw.matchAll(/\[([^\]]{2,2000})\]\s*TJ/gu), match => match[1] ?? ''));
@@ -998,16 +1200,166 @@ process.stdout.write(JSON.stringify({ pageCount: doc.numPages, pageLimit, text:
             .join('')
             .trim();
     }
+    toPdfTextItem(item) {
+        if (!item || typeof item !== 'object' || !('str' in item))
+            return undefined;
+        const record = item;
+        const transform = Array.isArray(record.transform) ? record.transform : [];
+        return {
+            str: String(record.str ?? ''),
+            x: Number(transform[4] ?? 0),
+            y: Number(transform[5] ?? 0),
+            width: Number(record.width ?? 0),
+            height: Number(record.height ?? Math.abs(Number(transform[3] ?? 0))),
+            fontName: typeof record.fontName === 'string' ? record.fontName : undefined,
+        };
+    }
+    layoutPdfTextItems(items, pageNumber) {
+        if (items.length === 0)
+            return '';
+        const rows = this.groupPdfItemsIntoRows(items);
+        const columnSplit = this.detectPdfColumnSplit(rows);
+        const orderedRows = columnSplit == null
+            ? rows.sort((a, b) => b.y - a.y || a.x - b.x)
+            : [
+                ...rows.filter(row => row.x < columnSplit).sort((a, b) => b.y - a.y || a.x - b.x),
+                ...rows.filter(row => row.x >= columnSplit).sort((a, b) => b.y - a.y || a.x - b.x),
+            ];
+        const markdown = this.rowsToPdfMarkdownWithTables(orderedRows);
+        return [`## PDF 第 ${pageNumber} 页`, markdown].join('\n\n');
+    }
+    groupPdfItemsIntoRows(items) {
+        const sorted = [...items].sort((a, b) => b.y - a.y || a.x - b.x);
+        const rows = [];
+        for (const item of sorted) {
+            const row = rows.find(candidate => Math.abs(candidate.y - item.y) <= Math.max(2, item.height * 0.55));
+            if (row)
+                row.items.push(item);
+            else
+                rows.push({ y: item.y, items: [item] });
+        }
+        return rows.map(row => {
+            const rowItems = row.items.sort((a, b) => a.x - b.x);
+            return {
+                text: rowItems.map(item => item.str.trim()).filter(Boolean).join(' ').replace(/\s+/gu, ' '),
+                x: Math.min(...rowItems.map(item => item.x)),
+                y: row.y,
+                height: Math.max(...rowItems.map(item => item.height || 0)),
+            };
+        }).filter(row => row.text);
+    }
+    detectPdfColumnSplit(rows) {
+        if (rows.length < 8)
+            return undefined;
+        const xs = rows.map(row => row.x).sort((a, b) => a - b);
+        const gaps = xs.slice(1).map((x, index) => ({ gap: x - xs[index], left: xs[index], right: x })).sort((a, b) => b.gap - a.gap);
+        const largest = gaps[0];
+        if (!largest || largest.gap < 80)
+            return undefined;
+        const leftCount = rows.filter(row => row.x <= largest.left).length;
+        const rightCount = rows.filter(row => row.x >= largest.right).length;
+        return leftCount >= 3 && rightCount >= 3 ? (largest.left + largest.right) / 2 : undefined;
+    }
+    rowsToPdfMarkdownWithTables(rows) {
+        const output = [];
+        let index = 0;
+        while (index < rows.length) {
+            const tableRows = [];
+            let cursor = index;
+            while (cursor < rows.length) {
+                const cells = this.splitLikelyTableRow(rows[cursor].text);
+                if (cells.length < 2)
+                    break;
+                tableRows.push(cells);
+                cursor += 1;
+            }
+            if (tableRows.length >= 2) {
+                output.push('### PDF 表格区域');
+                output.push(this.toMarkdownTable(tableRows[0], tableRows.slice(1)));
+                index = cursor;
+                continue;
+            }
+            output.push(this.pdfRowToMarkdown(rows[index].text, rows[index].height, index));
+            index += 1;
+        }
+        return output.join('\n');
+    }
+    splitLikelyTableRow(text) {
+        const byLargeSpaces = text.split(/\s{2,}/u).map(cell => cell.trim()).filter(Boolean);
+        if (byLargeSpaces.length >= 2)
+            return byLargeSpaces;
+        const byPipes = text.split('|').map(cell => cell.trim()).filter(Boolean);
+        return byPipes.length >= 2 ? byPipes : [];
+    }
+    pdfRowToMarkdown(text, height, index) {
+        if (index === 0 && text.length <= 100)
+            return `# ${text}`;
+        if (height >= 14 && text.length <= 120)
+            return `## ${text}`;
+        if (/^(第[一二三四五六七八九十\d]+[章节]|\d+(?:\.\d+)*\s+)/u.test(text) && text.length <= 120)
+            return `### ${text}`;
+        return text;
+    }
+    toMarkdownDocument(text) {
+        const lines = text.split(/\r?\n/u).map(line => line.trim()).filter(Boolean);
+        return lines.map((line, index) => {
+            if (/^#{1,6}\s/u.test(line) || /^\|/u.test(line))
+                return line;
+            if (line.length <= 80 && !/[。！？.!?]$/u.test(line)) {
+                if (index === 0)
+                    return `# ${line}`;
+                if (/^(第[一二三四五六七八九十\d]+[章节]|\d+(?:\.\d+)*\s+)/u.test(line))
+                    return `## ${line}`;
+                return `### ${line}`;
+            }
+            return line;
+        }).join('\n\n');
+    }
+    normalizeMarkdownHeadings(text) {
+        return text
+            .split(/\r?\n/u)
+            .map(line => {
+            const trimmed = line.trim();
+            if (!trimmed)
+                return '';
+            if (/^#{1,6}\s/u.test(trimmed) || /^\|/u.test(trimmed))
+                return trimmed;
+            if (trimmed.length <= 80 && /^(第[一二三四五六七八九十\d]+[章节]|\d+(?:\.\d+)*\s+)/u.test(trimmed))
+                return `## ${trimmed}`;
+            return trimmed;
+        })
+            .join('\n');
+    }
     extractSvg(file) {
         const raw = fs.readFileSync(file.absolutePath, 'utf8');
-        const texts = this.matchAll(raw, /<text\b[^>]*>([\s\S]*?)<\/text>/giu).map(value => this.stripXml(value)).slice(0, 300);
-        const titles = this.matchAll(raw, /<(?:title|desc)\b[^>]*>([\s\S]*?)<\/(?:title|desc)>/giu).map(value => this.stripXml(value)).slice(0, 100);
+        const nodes = this.extractSvgSemanticNodes(raw);
         return {
-            text: [this.metadataOnlyText(file), `SVG 标题/描述:\n${titles.join('\n')}`, `SVG 文本节点:\n${texts.join('\n')}`].join('\n'),
-            metadata: { extractionMode: 'svg_text_nodes', vectorizable: true, textNodeCount: texts.length, contentCoverage: 'svg_text_title_desc' },
+            text: [this.metadataOnlyText(file), 'SVG 层级语义节点:', ...nodes].join('\n'),
+            metadata: { extractionMode: 'svg_text_nodes', semanticExtractionMode: 'svg_semantic_tree_nodes', vectorizable: true, textNodeCount: nodes.length, contentCoverage: 'svg_hierarchical_text_title_desc' },
             warnings: [],
         };
     }
+    extractSvgSemanticNodes(raw) {
+        const nodes = [];
+        const stack = [];
+        const tokenPattern = /<\/?([A-Za-z_][\w:.-]*)\b([^>]*)>|([^<>]+)/gu;
+        for (const match of raw.matchAll(tokenPattern)) {
+            const tag = match[1];
+            const attrs = match[2] ?? '';
+            const text = match[3]?.replace(/\s+/gu, ' ').trim();
+            const token = match[0];
+            if (tag && token.startsWith('</'))
+                stack.pop();
+            else if (tag && !token.endsWith('/>')) {
+                const id = /\bid="([^"]+)"/u.exec(attrs)?.[1];
+                stack.push(id ? `${tag}#${id}` : tag);
+            }
+            else if (text && ['text', 'title', 'desc'].includes(stack.at(-1)?.split('#')[0] ?? '')) {
+                nodes.push(`${stack.join(' > ')}: ${this.stripXml(text)}`);
+            }
+        }
+        return nodes;
+    }
     isTextReadable(file) {
         if (file.category === 'code' || file.category === 'web')
             return true;
@@ -1042,13 +1394,61 @@ process.stdout.write(JSON.stringify({ pageCount: doc.numPages, pageLimit, text:
             return [`${prefix || 'value'}: ${String(value)}`];
         }
         if (Array.isArray(value)) {
-            return value.slice(0, 50).flatMap((item, index) => this.flattenJson(item, `${prefix}[${index}]`));
+            return value.flatMap((item, index) => this.flattenJson(item, `${prefix}[${index}]`));
         }
         if (typeof value === 'object') {
             return Object.entries(value).flatMap(([key, item]) => this.flattenJson(item, prefix ? `${prefix}.${key}` : key));
         }
         return [];
     }
+    atomicJsonObjects(value, prefix = '$') {
+        if (value == null || typeof value !== 'object')
+            return [];
+        if (Array.isArray(value))
+            return value.flatMap((item, index) => this.atomicJsonObjects(item, `${prefix}[${index}]`));
+        const entries = Object.entries(value);
+        const current = `${prefix}: ${JSON.stringify(value)}`;
+        const children = entries.flatMap(([key, item]) => this.atomicJsonObjects(item, `${prefix}.${key}`));
+        return [current, ...children];
+    }
+    flattenYamlByIndent(raw) {
+        const stack = [];
+        const lines = [];
+        for (const line of raw.split(/\r?\n/u)) {
+            if (!line.trim() || /^\s*#/u.test(line))
+                continue;
+            const match = /^(\s*)([-\w.]+)\s*:\s*(.*)$/u.exec(line);
+            if (!match)
+                continue;
+            const indent = match[1].length;
+            const key = match[2];
+            const value = match[3].trim();
+            while (stack.length > 0 && stack.at(-1).indent >= indent)
+                stack.pop();
+            const pathName = [...stack.map(item => item.key), key].join('.');
+            if (value)
+                lines.push(`${pathName}: ${value}`);
+            stack.push({ indent, key });
+        }
+        return lines;
+    }
+    flattenXmlPaths(raw) {
+        const lines = [];
+        const stack = [];
+        const tokenPattern = /<\/?([A-Za-z_][\w:.-]*)\b[^>]*>|([^<>]+)/gu;
+        for (const match of raw.matchAll(tokenPattern)) {
+            const tag = match[1];
+            const text = match[2]?.replace(/\s+/gu, ' ').trim();
+            const token = match[0];
+            if (tag && token.startsWith('</'))
+                stack.pop();
+            else if (tag && !token.endsWith('/>'))
+                stack.push(tag);
+            else if (text && stack.length > 0)
+                lines.push(`${stack.join('.')}: ${this.stripXml(text)}`);
+        }
+        return lines.filter(line => !line.endsWith(':'));
+    }
     metadataOnlyText(file) {
         const fileName = path.basename(file.relativePath);
         const directory = path.dirname(file.relativePath);