Update app.py

app.py

@@ -207,6 +207,22 @@ def ocr_zone(image_path, y_start_frac, y_end_frac):
     return ""
 
 
+def close_unclosed_html(md):
+    """Close any unclosed <table>, <tbody>, <tr> tags to prevent layout bleed."""
+    if not md:
+        return md
+    import re
+    open_tags = re.findall(r'<(table|tbody|thead|tr|td|th)[\s>]', md)
+    close_tags = re.findall(r'</(table|tbody|thead|tr|td|th)>', md)
+    # Count unclosed tags and close them in reverse order
+    for tag in reversed(['table', 'tbody', 'thead', 'tr', 'td', 'th']):
+        opened = open_tags.count(tag)
+        closed = close_tags.count(tag)
+        if opened > closed:
+            md += '</{}>'.format(tag) * (opened - closed)
+    return md
+
+
 def get_page_md_and_regions(page_result):
     md = ""
     if hasattr(page_result, "markdown_result") and page_result.markdown_result:
@@ -286,7 +302,7 @@ def run_ocr(uploaded_file):
                     parts.append(hdr.strip())
 
             if page_md:
-                parts.append(page_md)
+                parts.append(close_unclosed_html(page_md))
 
             # Footer extraction removed — MaaS API already captures all footer
             # content correctly inside page_md. Extracting again causes duplication.