enhancement: include coords in fast (#626)

Makes the bounding box coordinates available when using fast strategy.

* Refactored partition_text to make the workflow of categorizing an element purely from the text available without running the entirety of partition_text.
* Transformed the coordinates from pdf space into pixel space to be consistent with hi_res. We will probably want to revisit the coordinate system soon.

CHANGELOG.md

@@ -1,7 +1,9 @@
-## 0.6.9-dev0
+## 0.6.9-dev1
 
 ### Enhancements
 
+* fast strategy for pdf now keeps element bounding box data
+
 ### Features
 
 ### Fixes

test_unstructured/partition/test_pdf.py

@@ -404,7 +404,15 @@ 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 elements[0] == Title(
+        "eastern mediterranean",
+        coordinates=(
+            (193.1741, 71.94000000000005),
+            (193.1741, 91.94000000000005),
+            (418.6881, 91.94000000000005),
+            (418.6881, 71.94000000000005),
+        ),
+    )
 
     assert isinstance(elements[1], NarrativeText)
     assert str(elements[1]).startswith("We")
@@ -412,4 +420,10 @@ def test_partition_pdf_fast_groups_text_in_text_box():
 
     assert elements[3] == Title(
         "kilograms CO₂e/boe carbon intensity from our Eastern Mediterranean operations in 2022",
+        coordinates=(
+            (69.4871, 222.4357),
+            (69.4871, 272.1607),
+            (197.8209, 272.1607),
+            (197.8209, 222.4357),
+        ),
     )

unstructured/__version__.py

@@ -1 +1 @@
-__version__ = "0.6.9-dev0"  # pragma: no cover
+__version__ = "0.6.9-dev1"  # pragma: no cover

unstructured/partition/pdf.py

@@ -22,7 +22,7 @@ from unstructured.partition.common import (
     spooled_to_bytes_io_if_needed,
 )
 from unstructured.partition.strategies import determine_pdf_or_image_strategy
-from unstructured.partition.text import partition_text
+from unstructured.partition.text import element_from_text, partition_text
 from unstructured.utils import requires_dependencies
 
 
@@ -275,11 +275,16 @@ def _process_pdfminer_pages(
 
     for i, page in enumerate(extract_pages(fp)):  # type: ignore
         metadata = ElementMetadata(filename=filename, page_number=i + 1)
+        height = page.height
 
         text_segments = []
         for obj in page:
             # NOTE(robinson) - "Figure" is an example of an object type that does
             # not have a get_text method
+            x1, y2, x2, y1 = obj.bbox
+            y1 = height - y1
+            y2 = height - y2
+
             if not hasattr(obj, "get_text"):
                 continue
             _text = obj.get_text()
@@ -287,13 +292,10 @@ def _process_pdfminer_pages(
             _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
-            elements.append(element)
+                element = element_from_text(_text)
+                element.coordinates = ((x1, y1), (x1, y2), (x2, y2), (x2, y1))
+                element.metadata = metadata
+                elements.append(element)
 
         if include_page_breaks:
             elements.append(PageBreak())

unstructured/partition/text.py

@@ -90,17 +90,22 @@ def partition_text(
     for ctext in file_content:
         ctext = ctext.strip()
 
-        if ctext == "":
-            continue
-        if is_bulleted_text(ctext):
-            elements.append(ListItem(text=clean_bullets(ctext), metadata=metadata))
-        elif is_us_city_state_zip(ctext):
-            elements.append(Address(text=ctext, metadata=metadata))
-        elif is_possible_narrative_text(ctext):
-            elements.append(NarrativeText(text=ctext, metadata=metadata))
-        elif is_possible_title(ctext):
-            elements.append(Title(text=ctext, metadata=metadata))
-        else:
-            elements.append(Text(text=ctext, metadata=metadata))
+        if ctext:
+            element = element_from_text(ctext)
+            element.metadata = metadata
+            elements.append(element)
 
     return elements
+
+
+def element_from_text(text: str) -> Element:
+    if is_bulleted_text(text):
+        return ListItem(text=clean_bullets(text))
+    elif is_us_city_state_zip(text):
+        return Address(text=text)
+    elif is_possible_narrative_text(text):
+        return NarrativeText(text=text)
+    elif is_possible_title(text):
+        return Title(text=text)
+    else:
+        return Text(text=text)