fix: preserve multiline DeepSeek and Unlimited OCR blocks

docling/utils/deepseekocr_utils.py

@@ -158,23 +158,16 @@ def _collect_annotation_content(
                 break
             ii += 1
     else:
-        # Original logic for other labels
+        # Text and captions can span lines and paragraphs within one annotation.
         while i < len(lines):
-            content_line = lines[i].strip()
-            if content_line:
-                if re.match(annotation_pattern, content_line):
-                    break
-                visited_lines.add(i)
+            if re.match(annotation_pattern, lines[i].strip()):
+                break
+            if i not in visited_lines:
                 content_lines.append(lines[i].rstrip())
-                i += 1
-                if label_str not in ["figure", "image"]:
-                    break
-            else:
-                i += 1
-                if content_lines:
-                    break
+                visited_lines.add(i)
+            i += 1
 
-    return "\n".join(content_lines), i
+    return "\n".join(content_lines).strip("\n"), i
 
 
 def _process_annotation_item(

tests/test_ocr_multiline_blocks.py

@@ -0,0 +1,60 @@
+# SPDX-FileCopyrightText: The Docling Contributors
+# SPDX-License-Identifier: MIT
+
+import pytest
+from docling_core.types.doc import DocItemLabel, Size
+
+from docling.utils.deepseekocr_utils import (
+    parse_deepseekocr_markdown,
+    parse_unlimited_ocr_markdown,
+)
+
+
+@pytest.mark.parametrize("parser", ["deepseek", "unlimited"])
+def test_text_block_keeps_continuation_lines_and_following_table(parser: str) -> None:
+    def marker(label: str, top: int) -> str:
+        if parser == "deepseek":
+            return f"{label}[[100,{top},900,{top + 100}]]\n"
+        return f"<|det|>{label} [100,{top},900,{top + 100}]<|/det|>"
+
+    source = (
+        marker("text", 100)
+        + "First line\nSecond line\n\nThird paragraph\n"
+        + marker("table", 400)
+        + "<table><tr><td>Total</td></tr></table>\n"
+        + marker("table_caption", 600)
+        + "Caption first line\nCaption second line\n"
+        + marker("text", 800)
+        + "Final block"
+    )
+    parse = (
+        parse_deepseekocr_markdown
+        if parser == "deepseek"
+        else parse_unlimited_ocr_markdown
+    )
+    doc = parse(source, Size(width=600, height=800), page_no=2)
+    text = next(item for item in doc.texts if item.label == DocItemLabel.TEXT)
+    assert text.text == "First line\nSecond line\n\nThird paragraph"
+    assert text.prov[0].page_no == 2
+    assert len(doc.tables) == 1
+    assert doc.tables[0].data.grid[0][0].text == "Total"
+    caption = doc.tables[0].captions[0].resolve(doc)
+    assert caption.text == "Caption first line\nCaption second line"
+    assert doc.texts[-1].text == "Final block"
+    assert "Second line" in doc.export_to_markdown()
+
+
+def test_caption_does_not_claim_table_markup_already_collected() -> None:
+    source = (
+        "table[[100,100,900,300]]\n"
+        "table_caption[[100,400,900,500]]\nCaption first\nCaption second\n\n"
+        "<table><tr><td>Total</td></tr></table>\n\n"
+        "text[[100,600,900,700]]\nFinal block"
+    )
+    doc = parse_deepseekocr_markdown(source, Size(width=600, height=800), page_no=1)
+    assert doc.tables[0].data.grid[0][0].text == "Total"
+    assert (
+        doc.tables[0].captions[0].resolve(doc).text == "Caption first\nCaption second"
+    )
+    assert all("<table>" not in item.text for item in doc.texts)
+    assert doc.texts[-1].text == "Final block"