导出md文件，会把ocr原始结果中的识别出的内容遗失掉

[zjfchinespace]

<|ref|>sub_title<|/ref|><|det|>[[230, 65, 950, 135]]<|/det|>

光明师点窍一百一十一、四柱空亡对应的情况

<|ref|>sub_title<|/ref|><|det|>[[170, 157, 360, 186]]<|/det|>

1、年支空亡

<|ref|>text<|/ref|><|det|>[[157, 183, 535, 220], [157, 214, 535, 248], [159, 250, 406, 276], [159, 272, 536, 303], [159, 288, 884, 334], [161, 331, 510, 359]]<|/det|>
①表示不得祖荫庇护，祖上贫困
②不得父母帮助或父母能力不行
③自幼与父母缘分浅
④容易年少与父母分开或被寄养
⑤一生劳苦，需凭己力奋斗，无太多父母之力依靠或无贵人
⑥也代表祖上出天亡平庸之人

<|ref|>sub_title<|/ref|><|det|>[[177, 380, 366, 406]]<|/det|>

2、月支空亡

<|ref|>text<|/ref|><|det|>[[169, 410, 737, 439]]<|/det|>
①表示手足缘薄，也代表得不到手足朋友的帮助

<|ref|>text<|/ref|><|det|>[[170, 442, 650, 467]]<|/det|>
②代表兄弟姐妹稀少或有伤克，严重独生

<|ref|>text<|/ref|><|det|>[[171, 471, 570, 494]]<|/det|>
③与兄弟姐妹分较浅，感情不深

<|ref|>text<|/ref|><|det|>[[171, 498, 768, 522]]<|/det|>
④兄弟姐妹能力不强或对自己没有帮助，拖累自己

<|ref|>sub_title<|/ref|><|det|>[[184, 543, 370, 568]]<|/det|>

3、日支空亡

<|ref|>text<|/ref|><|det|>[[175, 580, 465, 602]]<|/det|>
①表示配偶无助或少助力

<|ref|>text<|/ref|><|det|>[[175, 608, 468, 630]]<|/det|>
②婚姻不稳定，感情不好

<|ref|>text<|/ref|><|det|>[[176, 635, 521, 659], [176, 660, 801, 700]]<|/det|>
③难有刻骨铭心的爱情或感情
④与配偶感情较淡，严重有破婚之象或配偶身体不适

<|ref|>text<|/ref|><|det|>[[177, 688, 580, 712]]<|/det|>
⑤容易感情不稳定，或婚外情

<|ref|>sub_title<|/ref|><|det|>[[189, 731, 377, 756]]<|/det|>

4、时支空亡

<|ref|>text<|/ref|><|det|>[[180, 766, 394, 790]]<|/det|>
①晚年不利，多病

<|ref|>text<|/ref|><|det|>[[180, 794, 499, 822], [180, 820, 660, 864]]<|/det|>
②代表与子女缘分薄，少聚
③子女不利，子女能力不强或对命主不好

上述是pdf识别后OCR原始结果。但是导出markdown文件后，结果丢失了一些。
下面就是导出的md内容：

光明师点窍一百一十一、四柱空亡对应的情况

1、年支空亡

2、月支空亡

①表示手足缘薄，也代表得不到手足朋友的帮助

②代表兄弟姐妹稀少或有伤克，严重独生

③与兄弟姐妹分较浅，感情不深

④兄弟姐妹能力不强或对自己没有帮助，拖累自己

3、日支空亡

①表示配偶无助或少助力

②婚姻不稳定，感情不好

⑤容易感情不稳定，或婚外情

4、时支空亡

①晚年不利，多病

“## 1、年支空亡”下面的6行文字没了，在“OCR原始结果”中是完全识别出来了，当导出就有些文字消失了

[zjfchinespace]

我用qwen code识别出了这个bug

✦ Summary

I've successfully fixed the markdown export issue where content was being lost. Here's what was wrong and what I fixed:

Root Cause
The bug was in /home/zjf/ai/DeepSeek-OCR-WebUI-main/frontend/src/services/doc-gen/markdown.ts in the processBlockMatch method.     
When OCR blocks contained multiple bounding boxes (like [[180, 794, 499, 822], [180, 820, 660, 864]]), the coordinate parser       
would split by comma and get 8 numbers instead of 4, causing the function to return null and silently discard the content.

This affected these blocks in your OCR output:
 - ②代表与子女缘分薄，少聚\n③子女不利，子女能力不强或对命主不好 (multi-box)
 - Several other blocks with multiple coordinate sets

The Fix
I added a new parseCoordinates() method that:
 1. Handles both single-box (x1, y1, x2, y2) and multi-box (x1, y1, x2, y2], [x1, y1, x2, y2) formats
 2. Extracts the first bounding box as the primary coordinate
 3. Properly validates and returns coordinates for all blocks

Verification
 - ✅ All 16 blocks from your OCR output are now parsed correctly
 - ✅ All existing unit tests pass (14/14)
 - ✅ Frontend builds successfully
 - ✅ The last two blocks that were previously lost are now included in the exported markdown

The fix is now live in the built frontend. When you run the application and export to markdown, all content will be preserved!     

markdown.ts
附件就是修改后的文件

[zjfchinespace]

Root Cause

The issue was in the collectOverlappingBlocks method in
～/ai/DeepSeek-OCR-WebUI-main/frontend/src/services/doc-gen/markdown.ts.

When grouping blocks into visual rows for multi-column layout, the algorithm would expand its search window too aggressively,      
pulling content from different semantic sections into the same table. Specifically:

 1. Text from section "一、从身体" (Y: 151-382) and image (Y: 160-465) were grouped together
 2. The window expanded to include Y: 151-465
 3. Text from section "二、财富" (Y: 410-500) overlapped with this range and got incorrectly grouped
 4. This caused the wealth section text to appear in a table with earlier content, leaving the "## 二、财富" subtitle without its  
    text content

The Fix

Added a boundary check in collectOverlappingBlocks to prevent grouping blocks across heading (sub_title/title) boundaries:

 - When considering adding a block to a visual row, check if it comes after the seed block
 - If yes, check if there's a heading between them in the Y-axis range
 - If a heading exists between them, don't group them (they belong to different sections)

Files Modified

 - ～/ai/DeepSeek-OCR-WebUI-main/frontend/src/services/doc-gen/markdown.ts
   - Modified collectOverlappingBlocks method
   - Added hasHeadingInYRange helper method

Verification

✓ Build succeeds without errors
✓ All content now appears in correct sections
✓ Section boundaries are respected during layout analysis

The fix ensures that:
 - All text content appears in the exported markdown
 - Content stays under its proper section heading
 - Multi-column layouts only include content that truly belongs together

markdown.ts