fix: OCR scanned pages when mixed with text pages

Author: Brumbelow

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:

packages/markitdown-ocr/tests/ocr_test_data/pdf_text_then_blank.pdf

@@ -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)
 # ---------------------------------------------------------------------------