From 83f75895b3eb6bf8616b310fe162091eb93c6d9c Mon Sep 17 00:00:00 2001
From: Brumbelow <andrew.brumbelow@gmail.com>
Date: Wed, 22 Apr 2026 13:38:53 -0400
Subject: [PATCH] fix: OCR scanned pages when mixed with text pages

---
 .../markitdown_ocr/_pdf_converter_with_ocr.py |  28 ++++++++++++++++++
 .../ocr_test_data/pdf_text_then_blank.pdf     | Bin 0 -> 1762 bytes
 .../tests/test_pdf_converter.py               |  16 ++++++++++
 3 files changed, 44 insertions(+)
 create mode 100644 packages/markitdown-ocr/tests/ocr_test_data/pdf_text_then_blank.pdf

diff --git a/packages/markitdown-ocr/src/markitdown_ocr/_pdf_converter_with_ocr.py b/packages/markitdown-ocr/src/markitdown_ocr/_pdf_converter_with_ocr.py
index c1dc0f613..e79582e67 100644
--- a/packages/markitdown-ocr/src/markitdown_ocr/_pdf_converter_with_ocr.py
+++ b/packages/markitdown-ocr/src/markitdown_ocr/_pdf_converter_with_ocr.py
@@ -280,6 +280,13 @@ def convert(
                             text_content = page.extract_text() or ""
                             if text_content.strip():
                                 markdown_content.append(text_content.strip())
+                            else:
+                                # Scanned page with no image objects or text layer:
+                                # render the whole page and OCR it, so it doesn't
+                                # silently vanish from the output.
+                                page_ocr = self._ocr_page(page, ocr_service)
+                                if page_ocr:
+                                    markdown_content.append(page_ocr)
                     else:
                         # No OCR, just extract text
                         text_content = page.extract_text() or ""
@@ -337,6 +344,27 @@ def _extract_page_images(self, pdf_bytes: io.BytesIO, page_num: int) -> list[dic
 
         return images
 
+    def _ocr_page(self, page: Any, ocr_service: LLMVisionOCRService) -> str:
+        """
+        Render a single pdfplumber page to PNG and run OCR on it.
+
+        Returns a formatted OCR block, a 'no text' marker if OCR is empty,
+        or an error marker if rendering/OCR raises.
+        """
+        try:
+            page_img = page.to_image(resolution=300)
+            img_stream = io.BytesIO()
+            page_img.original.save(img_stream, format="PNG")
+            img_stream.seek(0)
+
+            ocr_result = ocr_service.extract_text(img_stream)
+            text = (ocr_result.text or "").strip()
+            if text:
+                return f"*[Image OCR]\n{text}\n[End OCR]*"
+            return "*[No text could be extracted from this page]*"
+        except Exception as e:
+            return f"*[Error processing page {page.page_number}: {str(e)}]*"
+
     def _ocr_full_pages(
         self, pdf_bytes: io.BytesIO, ocr_service: LLMVisionOCRService
     ) -> str:
diff --git a/packages/markitdown-ocr/tests/ocr_test_data/pdf_text_then_blank.pdf b/packages/markitdown-ocr/tests/ocr_test_data/pdf_text_then_blank.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..c64d896b6ec29b4cd78fa935fed49408ccdf234a
GIT binary patch
literal 1762
zcmdT_TX*6%5PtWsm_1oeNH<{1##}a(Znz~hBwPZUwm`bFEx=ijd}Mi>^r3&Ced~LF
zYe!rXm*$*3r*HLxr5TMz^Nqe4W%|u_j^rI9^XDJG{{=%938}i&gA2x)q>2TA77z$O
zSu?I6D<Z~Ykw~AN7z8X3-h)xC8fKe-jpER#)eOc1t=95>xlUu&7U;mdWUDVsg+2wd
z#eETkd;#WN$j^Bkz9??!nt^s|-X@-UiclEgA7wq@3i?VN<)n^k;1Y%!Ls`cI<~0jK
zS{IKnLD_;Lv7uD4VFqTON{y+swHvZ6t{66Wbg&vAdgx2x-!g>`Oq~pK#2yhyXH6HZ
zz8f36u>&&<3(TQVm^|EWjwIz9ms)SY9zY*ArvMbtbf2qd0F~|W{|f+a3V_CTEdsoz
zam;Q6$B9<O`%G`Iut61FiA5?2#gvcpL@kAcEaieftwj=_fZ33Ys!;HzPHndN-nQ(b
zRj_U15ZkhHR^iyPj`wzlo3Gf@FG2<}WYGen`<fHH^dRvWhVW7p`u4k_C#igb986Qi
zFg-chTrRTpf%#Y`36CS@hx3pHc-|;fD-1c|N>4T#Nuvm>%WTtCoMwV%;|mjT2iBE&
z6=L8M&@>-{=E_j>cbmseYExO%wBI{uhbx5;khtf!jZUZRHW`+juAdd=a3`RrF5?Te
z1mf7*QLHf0YvY1OA6gdu)Ju-LllM*D^D^})xNaUV<eWqw-Yv(%jRu8cFaNPMc>6$_
zUTJpLwY^*+Z_dW!Z!*p#f7TndnycDfH>3Rd(e$TkuPc}HEHCBG!``QIvM$wIchSd-
zFyD#T>G-Jn{SR;AceBs!V6wcd{#L_QvWsTdRCG=JM|5uGpWp^dtMzUB)EZAOozXP9
z`EhhVxcdB^@@t*`>Dk@UnB3b=xa_yCfA06+_TD)6rww=TN=%QE%?%5PWmwN26kXSG
zVGhr_B6bt~0e{h9A3D^L3PL=vb6&A?%UE&;c10H%a-gda<jbM1^sh6pf>QA*Ny8PB
zo6g>8f-$KwuUsyZU`~S)aV^&?lq((y9M2_$76MYKecfVaWKs{z;`Ik6mxCb0UUBf$
q7-omh1?UbjG_bo01-j$l{5JfR>C|s{OQqUe25|}%Ba><M+Q#1!?(b~?

literal 0
HcmV?d00001

diff --git a/packages/markitdown-ocr/tests/test_pdf_converter.py b/packages/markitdown-ocr/tests/test_pdf_converter.py
index 5d4adcc5e..5849c0323 100644
--- a/packages/markitdown-ocr/tests/test_pdf_converter.py
+++ b/packages/markitdown-ocr/tests/test_pdf_converter.py
@@ -189,6 +189,22 @@ def test_pdf_scanned_report(svc: MockOCRService) -> None:
     assert _convert("pdf_scanned_report.pdf", svc) == expected
 
 
+# ---------------------------------------------------------------------------
+# Mixed text + scanned pages: a page with neither image objects nor a text
+# layer must still be OCR'd inline rather than silently dropped.
+# Regression test for microsoft/markitdown#1791.
+# ---------------------------------------------------------------------------
+
+
+def test_pdf_text_then_blank_page_is_ocred(svc: MockOCRService) -> None:
+    expected = (
+        "## Page 1\n\n\n"
+        "First page content, extracted as text.\n\n\n"
+        f"## Page 2\n\n\n{_OCR_BLOCK}"
+    )
+    assert _convert("pdf_text_then_blank.pdf", svc) == expected
+
+
 # ---------------------------------------------------------------------------
 # Scanned PDF fallback path (pdfplumber finds no text → full-page OCR)
 # ---------------------------------------------------------------------------