From 1f6aacf7444c7f1339c2e0b2a0d1364fb626226c Mon Sep 17 00:00:00 2001
From: Yi-111-a <153097222+Yi-111-a@users.noreply.github.com>
Date: Tue, 20 Oct 2026 07:44:12 +0800
Subject: [PATCH] fix(deepseek-ocr): keep continuation lines in non-table
 blocks

_collect_annotation_content stopped after the first non-empty line for
labels other than tables and pictures, so a multi-line text/title block
lost every line after the first one. Unlimited OCR delegates to the same
collector after normalizing its markers, so it lost the same lines.

Now every non-table block collects consecutive non-empty lines until the
next annotation or the first blank line, so all content in the first
block survives until the next annotation. Empty lines, following
annotations, captions, and table markup keep their boundaries.

Fixes #4905

Signed-off-by: Yi-111-a <153097222+Yi-111-a@users.noreply.github.com>
---
 docling/utils/deepseekocr_utils.py   |  9 ++---
 tests/test_unlimited_ocr_markdown.py | 55 ++++++++++++++++++++++++++++
 2 files changed, 59 insertions(+), 5 deletions(-)

diff --git a/docling/utils/deepseekocr_utils.py b/docling/utils/deepseekocr_utils.py
index 76a39aaca5..f7d8161796 100644
--- a/docling/utils/deepseekocr_utils.py
+++ b/docling/utils/deepseekocr_utils.py
@@ -158,17 +158,16 @@ def _collect_annotation_content(
                 break
             ii += 1
     else:
-        # Original logic for other labels
+        # Collect the block's content lines: consecutive non-empty lines up to the
+        # next annotation or the first blank line, so continuation lines survive.
         while i < len(lines):
             content_line = lines[i].strip()
+            if re.match(annotation_pattern, content_line):
+                break
             if content_line:
-                if re.match(annotation_pattern, content_line):
-                    break
                 visited_lines.add(i)
                 content_lines.append(lines[i].rstrip())
                 i += 1
-                if label_str not in ["figure", "image"]:
-                    break
             else:
                 i += 1
                 if content_lines:
diff --git a/tests/test_unlimited_ocr_markdown.py b/tests/test_unlimited_ocr_markdown.py
index dd0d44da46..4cabf21c15 100644
--- a/tests/test_unlimited_ocr_markdown.py
+++ b/tests/test_unlimited_ocr_markdown.py
@@ -71,3 +71,58 @@ def test_parse_builds_document_with_heading_and_table():
 
 def test_response_format_enum_value():
     assert ResponseFormat.UNLIMITED_OCR_MARKDOWN.value == "unlimited_ocr_markdown"
+
+
+def test_continuation_lines_survive_in_a_text_block():
+    raw = (
+        "text[[100,100,900,300]]\n"
+        "First line\n"
+        "Second line\n"
+        "\n"
+        "text[[100,400,900,500]]\n"
+        "Next block\n"
+    )
+
+    doc = parse_unlimited_ocr_markdown(raw, PAGE_SIZE, page_no=1, filename="page.png")
+    markdown = doc.export_to_markdown()
+
+    assert "First line" in markdown
+    assert "Second line" in markdown
+    assert "Next block" in markdown
+    texts = [t.text for t in doc.texts]
+    assert any("First line" in t and "Second line" in t for t in texts)
+
+
+def test_unlimited_ocr_continuation_lines_survive_in_a_text_block():
+    raw = (
+        "<|det|>text [100, 100, 900, 300]<|/det|>First line\n"
+        "Second line\n"
+        "\n"
+        "<|det|>text [100, 400, 900, 500]<|/det|>Next block\n"
+    )
+
+    doc = parse_unlimited_ocr_markdown(raw, PAGE_SIZE, page_no=1, filename="page.png")
+    texts = [t.text for t in doc.texts]
+
+    assert len(texts) >= 2
+    assert any("First line" in t and "Second line" in t for t in texts)
+
+
+def test_annotation_and_blank_line_still_end_a_block():
+    raw = (
+        "text[[100,100,900,300]]\n"
+        "First line\n"
+        "Second line\n"
+        "\n"
+        "\n"
+        "Loose trailing text\n"
+        "text[[100,400,900,500]]\n"
+        "Next block\n"
+    )
+
+    doc = parse_unlimited_ocr_markdown(raw, PAGE_SIZE, page_no=1, filename="page.png")
+    texts = [t.text for t in doc.texts]
+
+    assert any(t == "First line\nSecond line" for t in texts)
+    assert any("Next block" in t for t in texts)
+    assert not any("Loose trailing text" in t for t in texts)