In the case of an RTL language being extracted via pdfminer.six, fall back to forced OCR, which handles RTL text better

src/paperless_tesseract/parsers.py

@@ -13,6 +13,10 @@ class NoTextFoundException(Exception):
     pass
 
 
+class RtlLanguageException(Exception):
+    pass
+
+
 class RasterisedDocumentParser(DocumentParser):
     """
     This parser uses Tesseract to try and get some text out of a rasterised
@@ -125,7 +129,26 @@ class RasterisedDocumentParser(DocumentParser):
             stripped = post_process_text(pdfminer_extract_text(pdf_file))
 
             self.log("debug", f"Extracted text from PDF file {pdf_file}")
+
+            # pdfminer.six does not handle RTL text
+            # as a hack, for some languages, return no text, to force
+            # OCRMyPdf/Tesseract do handle this correctly
+            from langdetect import detect
+
+            lang = detect(stripped)
+
+            self.log("debug", f"Detected language {lang}")
+
+            if lang in {
+                "ar",  # Arabic
+                "he",  # Hebrew,
+                "fa",  # Persian
+            }:
+                raise RtlLanguageException()
             return stripped
+        except RtlLanguageException:
+            self.log("warning", f"Detected RTL language {lang}")
+            return None
         except Exception:
             # TODO catch all for various issues with PDFminer.six.
             #  If PDFminer fails, fall back to OCR.
@@ -305,7 +328,7 @@ class RasterisedDocumentParser(DocumentParser):
             )
             if original_has_text:
                 self.text = text_original
-        except (NoTextFoundException, InputFileError) as e:
+        except (NoTextFoundException, RtlLanguageException, InputFileError) as e:
             self.log(
                 "warning",
                 f"Encountered an error while running OCR: {str(e)}. "

src/paperless_tesseract/tests/test_parser.py

@@ -588,6 +588,39 @@ class TestParser(DirectoriesMixin, TestCase):
             params = parser.construct_ocrmypdf_parameters("", "", "", "")
             self.assertNotIn("deskew", params)
 
+    def test_rtl_language_detection(self):
+        """
+        GIVEN:
+            - File with text in an RTL language
+        WHEN:
+            - Document is parsed
+        THEN:
+            - Text from the document is extracted
+        """
+        parser = RasterisedDocumentParser(None)
+        with mock.patch.object(
+            parser,
+            "construct_ocrmypdf_parameters",
+            wraps=parser.construct_ocrmypdf_parameters,
+        ) as wrapped:
+
+            parser.parse(
+                os.path.join(self.SAMPLE_FILES, "rtl-test.pdf"),
+                "application/pdf",
+            )
+
+            # There isn't a good way to actually check this working, with RTL correctly return
+            #  as it would require tesseract-ocr-ara installed for everyone running the
+            #  test suite.  This test does provide the coverage though and attempts to ensure
+            # the force OCR happens
+            self.assertIsNotNone(parser.get_text())
+
+            self.assertEqual(parser.construct_ocrmypdf_parameters.call_count, 2)
+            # Check the last call kwargs
+            self.assertTrue(
+                parser.construct_ocrmypdf_parameters.call_args.kwargs["safe_fallback"],
+            )
+
 
 class TestParserFileTypes(DirectoriesMixin, TestCase):