fix: group together text from the same bounding box in partition_pdf with fast strategy (#542)

* switch to using PDF objects

* linting, linting, linting

* couple more tweaks

* added test for chevron-page

* version and changelog

* linting, linting, linting

* now processing 4 files

CHANGELOG.md

@@ -1,4 +1,4 @@
-## 0.6.3-dev0
+## 0.6.3-dev1
 
 ### Enhancements
 
@@ -10,6 +10,9 @@
 
 ### Fixes
 
+* Updates the grouping logic in the `partition_pdf` fast strategy to group together text
+  in the same bounding box.
+
 
 ## 0.6.2
 

test_unstructured/partition/test_pdf.py

@@ -319,3 +319,18 @@ def test_partition_pdf_fails_if_pdf_not_processable(
 
     with pytest.raises(ValueError):
         pdf.partition_pdf(filename=filename)
+
+
+def test_partition_pdf_fast_groups_text_in_text_box():
+    filename = os.path.join("example-docs", "chevron-page.pdf")
+    elements = pdf.partition_pdf(filename=filename, strategy="fast")
+
+    assert elements[0] == Title("eastern mediterranean")
+
+    assert isinstance(elements[1], NarrativeText)
+    assert str(elements[1]).startswith("We")
+    assert str(elements[1]).endswith("Jordan and Egypt.")
+
+    assert elements[3] == Title(
+        "kilograms CO₂e/boe carbon intensity from our Eastern Mediterranean operations in 2022",
+    )

test_unstructured_ingest/test-ingest-against-api.sh

@@ -15,8 +15,8 @@ PYTHONPATH=. ./unstructured/ingest/main.py \
 
 set +e
 
-if [ "$(find 'api-ingest-output' -type f -printf '.' | wc -c)" != 3 ]; then
+if [ "$(find 'api-ingest-output' -type f -printf '.' | wc -c)" != 4 ]; then
    echo
-   echo "3 files should have been created."
+   echo "4 files should have been created."
    exit 1
 fi

unstructured/__version__.py

@@ -1 +1 @@
-__version__ = "0.6.3-dev0"  # pragma: no cover
+__version__ = "0.6.3-dev1"  # pragma: no cover

unstructured/partition/pdf.py

@@ -1,12 +1,15 @@
+import re
 import warnings
-from io import StringIO
 from typing import BinaryIO, List, Optional, cast
 
+from pdfminer.high_level import extract_pages
 from pdfminer.pdfpage import PDFPage, PDFTextExtractionNotAllowed
 from pdfminer.utils import open_filename
 
+from unstructured.cleaners.core import clean_extra_whitespace
 from unstructured.documents.elements import Element, ElementMetadata, PageBreak
 from unstructured.logger import logger
+from unstructured.nlp.patterns import PARAGRAPH_PATTERN
 from unstructured.partition import _partition_via_api
 from unstructured.partition.common import (
     add_element_metadata,
@@ -285,27 +288,25 @@ def _process_pdfminer_pages(
     include_page_breaks: bool = False,
 ):
     """Uses PDF miner to split a document into pages and process them."""
-    from pdfminer.converter import TextConverter
-    from pdfminer.layout import LAParams
-    from pdfminer.pdfinterp import PDFPageInterpreter, PDFResourceManager
-
-    rsrcmgr = PDFResourceManager(caching=False)
-    laparams = LAParams()
-
     elements: List[Element] = []
 
-    for i, page in enumerate(PDFPage.get_pages(fp, check_extractable=True)):
+    for i, page in enumerate(extract_pages(fp)):  # type: ignore
         metadata = ElementMetadata(filename=filename, page_number=i + 1)
-        with StringIO() as output_string:
+
-            device = TextConverter(
+        text_segments = []
-                rsrcmgr,
+        for obj in page:
-                output_string,
+            # NOTE(robinson) - "Figure" is an example of an object type that does
-                codec=encoding,
+            # not have a get_text method
-                laparams=laparams,
+            if not hasattr(obj, "get_text"):
-            )
+                continue
-            interpreter = PDFPageInterpreter(rsrcmgr, device)
+            _text = obj.get_text()
-            interpreter.process_page(page)
+            _text = re.sub(PARAGRAPH_PATTERN, " ", _text)
-            text = output_string.getvalue()
+            _text = clean_extra_whitespace(_text)
+            if _text.strip():
+                text_segments.append(_text)
+
+        text = "\n\n".join(text_segments)
+
         _elements = partition_text(text=text)
         for element in _elements:
             element.metadata = metadata