diff --git a/packages/paper-qa-docling/src/paperqa_docling/reader.py b/packages/paper-qa-docling/src/paperqa_docling/reader.py
index fa3846c61..311d2d2aa 100644
--- a/packages/paper-qa-docling/src/paperqa_docling/reader.py
+++ b/packages/paper-qa-docling/src/paperqa_docling/reader.py
@@ -18,6 +18,7 @@
from docling_core.types.doc import (
DescriptionAnnotation,
DocItem,
+ FormulaItem,
PictureItem,
TableItem,
TextItem,
@@ -123,10 +124,14 @@ def parse_pdf_to_pages( # noqa: PLR0912
# NOTE: docling pages are 1-indexed
page_nums = [prov.page_no for prov in item.prov]
- if isinstance(item, TextItem):
+ if isinstance(item, TextItem | FormulaItem): # Handle items with text
+ item_text = item.text
+ if not item_text and isinstance(item, FormulaItem) and item.orig:
+ # Sometimes the sanitization of formula text fails, so use the original
+ item_text = item.orig
for page_num in page_nums:
new_text = (
- item.text if not content[str(page_num)][0] else "\n\n" + item.text
+ item_text if not content[str(page_num)][0] else "\n\n" + item_text
)
total_length += len(new_text)
if page_size_limit and total_length > page_size_limit:
@@ -137,7 +142,9 @@ def parse_pdf_to_pages( # noqa: PLR0912
)
content[str(page_num)][0] += new_text
- if parse_media and isinstance(item, PictureItem): # Handle images
+ if parse_media and isinstance( # Handle images and formulae
+ item, PictureItem | FormulaItem
+ ):
image_data = item.get_image(doc)
if image_data:
try:
@@ -154,14 +161,16 @@ def parse_pdf_to_pages( # noqa: PLR0912
img_bytes.seek(0) # Reset pointer before read to avoid empty data
media_metadata = {
- "type": "picture",
+ "type": "formula" if isinstance(item, FormulaItem) else "picture",
"width": image_data.width,
"height": image_data.height,
"bbox": item.prov[0].bbox.as_tuple(),
"images_scale": pipeline_options.images_scale,
}
annotations = [
- x for x in item.annotations if isinstance(x, DescriptionAnnotation)
+ x
+ for x in getattr(item, "annotations", [])
+ if isinstance(x, DescriptionAnnotation)
]
if len(annotations) == 1:
# We don't set this text in ParsedMedia.text because it's
diff --git a/packages/paper-qa-docling/tests/test_paperqa_docling.py b/packages/paper-qa-docling/tests/test_paperqa_docling.py
index 75e286e08..c0d327096 100644
--- a/packages/paper-qa-docling/tests/test_paperqa_docling.py
+++ b/packages/paper-qa-docling/tests/test_paperqa_docling.py
@@ -199,12 +199,12 @@ def test_media_deduplication() -> None:
all_media = [m for _, media in parsed_text.content.values() for m in media] # type: ignore[misc]
all_images = [m for m in all_media if m.info.get("type") == "picture"]
- # We allow for one table to be misinterpreted as an image
+ # We allow for one table to be misinterpreted as an image, and one logo to be missed
assert (
- 10 <= len(all_images) <= 11
+ 3 * 5 - 1 <= len(all_images) <= 3 * 5 + 1
), "Expected each image (one/page) and equation (one/page) to be read"
assert (
- len({m for m in all_images if cast(int, m.info["page_num"]) > 1}) <= 2
+ len({m for m in all_images if cast(int, m.info["page_num"]) > 1}) <= 3
), "Expected images/equations on all pages beyond 1 to be deduplicated"
all_tables = [m for m in all_media if m.info.get("type") == "table"]
diff --git a/tests/duplicate_media_template.md b/tests/duplicate_media_template.md
index 0c9f5fc87..4ed13834a 100644
--- a/tests/duplicate_media_template.md
+++ b/tests/duplicate_media_template.md
@@ -3,10 +3,14 @@
[//]: # "To generate `stub_data/duplicate_media.pdf` from this:"
-[//]: # "1. `pandoc duplicate_media_template.md --standalone --self-contained --katex -t html -o temp.html`"
+[//]: # "1. `pandoc duplicate_media_template.md --standalone --katex -t html -o temp.html`"
[//]: # "2. `Chromium --headless --disable-gpu --print-to-pdf=stub_data/duplicate_media.pdf --no-pdf-header-footer temp.html`"
[//]: # "3. `rm temp.html`"
+
+
+
Text under image 1.
@@ -28,6 +32,10 @@ $$
+
+
+
+
+
+
+
+
+