ValueError: min() arg is an empty sequence

[robvandijk]

Versions

• pymupdf==1.26.6
• pymupdf-layout==1.26.6
• pymupdf4llm==0.2.0

Steps to reproduce

import pymupdf.layout
import pymupdf
import pymupdf4llm

doc = pymupdf.open('file.pdf')
md = pymupdf4llm.to_markdown(doc)

Error encountered

Traceback (most recent call last):
  File "/opt/ori/lab/show.py", line 7, in <module>
    md = pymupdf4llm.to_markdown(doc)
  File "/usr/local/lib/python3.10/site-packages/pymupdf4llm/__init__.py", line 71, in to_markdown
    parsed_doc = parse_document(
  File "/usr/local/lib/python3.10/site-packages/pymupdf4llm/__init__.py", line 30, in parse_document
    return DL.parse_document(
  File "/usr/local/lib/python3.10/site-packages/pymupdf4llm/helpers/document_layout.py", line 716, in parse_document
    page.layout_information = utils.find_reading_order(
  File "/usr/local/lib/python3.10/site-packages/pymupdf4llm/helpers/utils.py", line 409, in find_reading_order
    min(b[0] for b in body_boxes),
ValueError: min() arg is an empty sequence

Notes

The pdf referred to above is attached.

file.pdf

[rahul-gj]

Sharing full stack-trace:

ValueError                                Traceback (most recent call last)
Cell In[1], line 8
      5 # download pdf from url: https://www.ema.europa.eu/en/documents/product-information/actrapid-epar-product-information_en.pdf
      7 doc = pymupdf.open("actrapid-epar-product-information_en.pdf")
----> 8 md = pymupdf4llm.to_markdown(doc)

File .venv\Lib\site-packages\pymupdf4llm\__init__.py:71, in to_markdown(doc, header, footer, pages, hdr_info, write_images, embed_images, ignore_images, ignore_graphics, detect_bg_color, image_path, image_format, image_size_limit, filename, force_text, page_chunks, page_separators, margins, dpi, page_width, page_height, table_strategy, graphics_limit, fontsize_limit, ignore_code, extract_words, show_progress, use_glyphs, ignore_alpha)
     39 def to_markdown(
     40     doc,
     41     *,
   (...)     69     ignore_alpha=False,
     70 ):
---> 71     parsed_doc = parse_document(
     72         doc,
     73         filename=filename,
     74         image_dpi=dpi,
     75         image_format=image_format,
     76         image_path=image_path,
     77         pages=pages,
     78     )
     79     return parsed_doc.to_markdown(
     80         header=header,
     81         footer=footer,
   (...)     84         ignore_code=ignore_code,
     85     )

File .venv\Lib\site-packages\pymupdf4llm\__init__.py:30, in parse_document(doc, filename, image_dpi, image_format, image_path, pages)
     22 def parse_document(
     23     doc,
     24     filename="",
   (...)     28     pages=None,
     29 ):
---> 30     return DL.parse_document(
     31         doc,
     32         filename=filename,
     33         image_dpi=image_dpi,
     34         image_format=image_format,
     35         image_path=image_path,
     36         pages=pages,
     37     )

File .venv\Lib\site-packages\pymupdf4llm\helpers\document_layout.py:716, in parse_document(doc, filename, image_dpi, image_format, image_path, pages)
    714 utils.add_image_orphans(page, blocks)
    715 utils.clean_tables(page, blocks)
--> 716 page.layout_information = utils.find_reading_order(
    717     page.rect, blocks, page.layout_information
    718 )
    720 # identify vector graphics to help find tables
    721 all_lines, all_boxes = utils.complete_table_structure(page)

File .venv\Lib\site-packages\pymupdf4llm\helpers\utils.py:409, in find_reading_order(page_rect, blocks, boxes, vertical_gap)
    405         body_boxes.append(box)
    407 # compute joined boxes of body
    408 joined_boxes = pymupdf.Rect(
--> 409     min(b[0] for b in body_boxes),
    410     min(b[1] for b in body_boxes),
    411     max(b[2] for b in body_boxes),
    412     max(b[3] for b in body_boxes),
    413 )
    415 # extract vectors contained in the TextPage
    416 min_bbox_height = min(b[3] - b[1] for b in body_boxes)

ValueError: min() iterable argument is empty

[JorjMcKie]

Ok, thanks for submitting this.
It is a true bug.

[rahul-gj]

in helpers/utils.py, line 409 where error is occuring, when I make this update, it goes away. I am no expert in pdf layout detection and this is just an observation and not solution.
old:

    joined_boxes = pymupdf.Rect(
        min(b[0] for b in body_boxes),
        min(b[1] for b in body_boxes),
        max(b[2] for b in body_boxes),
        max(b[3] for b in body_boxes),
    )

new:

joined_boxes = pymupdf.Rect(
        min((b[0] for b in body_boxes), default=0),
        min((b[1] for b in body_boxes), default=0),
        max((b[2] for b in body_boxes), default=0),
        max((b[3] for b in body_boxes), default=0),
    )```


and on line 416.
old: `min(b[3] - b[1] for b in body_boxes)`
new: `min((b[3] - b[1] for b in body_boxes), default=0)`

[JorjMcKie]

Thank you - I have a fix, easy case.

[JorjMcKie]

There now surfaces a performance issue because of extreme vector graphic cases on page 7, 10, 30 ...

So I need to check how to intercept them.

[robvandijk]

Thanks for looking into this. I have many more pdfs that are very slow to convert. Unfortunately I don't control the sources where they come from...

[JorjMcKie]

Analysis results:
There a handful of pages with 20K to 46K images!!! Yes: 46,000 tiny images of 3 x 3 points in size or less!
I will prepare a fix that ignores such nonsense. Of course, determining these fact themselves also takes time, but at least the overall duration goes down to 1.6 seconds per page.

[JorjMcKie]

BTW those zillions of micro-images are so-called "inline images" in the page /Contents. You don't see them when you do page.get_images().

[robvandijk]

Great, and I suspect that this will have a great performance impact on most of the other pdfs as well, if not all.

[jamie-lemon]

Why would someone put 46,000 tiny images of 3 x 3 into a PDF?!?

[JorjMcKie]

It is almost certainly a some-thing. It is something that look like a coarsely rasterized picture. Probably every single pixel was defined as an image inside the page's appearance stream.
Didn't bother to verify this hypothesis - not worth the effort.

[henry867]

I also had a case of significant performance issue during conversion with extremely dense tables inside, which takes over 50 mins to finish. Is it possible to fix it as well?

dense_table.pdf

[JorjMcKie]

I also had a case of significant performance issue during conversion with extremely dense tables inside, which takes over 50 mins to finish. Is it possible to fix it as well?

dense_table.pdf

@henry867 This looks like an entirely different case: no large images at all in that file.
Please submit a separate issue.