PowerPoint Export

[JorjMcKie]

Blaenau Gwent LAEP - Technical Report.pdf

[JorjMcKie]

As mentioned on Discord, documents with a highly complex layout like the above require radically simplifying assumptions to become reasonably accessible.

[JorjMcKie]

I am experimenting with additional parameters ignore_images/ignore_graphics of to_markdown.
In this case, the script

import pymupdf, pymupdf4llm, pathlib

filename = "test.pdf"
doc = pymupdf.open(filename)
md = pymupdf4llm.to_markdown(
    doc,
    # write_images=True,
    force_text=True,
    show_progress=True,
    ignore_images=True,
    ignore_graphics=True,
)
pathlib.Path(doc.name).with_suffix(".md").write_text(md)

delivers this output which is probably close to the desired one:
test.pdf.md

[JorjMcKie]

Among the remaining desirables are undue header classifications, which for example cause nonsense headers on page 1 and 2.

[dr-tariq-n-ahmad]

I am going to nick this code and try it. I will revert back with results. Thanks.

[JorjMcKie]

Wait! You would need my intermediate package changes. You can try with this zipped wheel:
pymupdf4llm-0.0.20-py3-none-any.zip
Install via python -m pip install --force-reinstall pymupdf4llm-0.0.20-py3-none-any.whl.

[JorjMcKie]

This file, pars pro toto, is a good example to see that probably every layout analysis is bound to fail at some point:
Just looking at the logo-like stuff at the top right corners: It is mixture of text, images and vector graphics. How for heavens sake could this ever be correctly interpreted as something to be consistently ignored as a whole?
Short of using advanced AI at least.

[JorjMcKie]

As you can see, at least some of the tables come out quite nicely.

If you have read the package's API documentation, then you know that its header identification is also quite simple. In essence it is based on classifying the full document's text based on a statistical analysis of font size frequencies:
The most frequent font size (and all smaller sizes) becomes "body text". Larger sizes are mapped to the HTML's header tags ("h1" through "h6"). No looking at bold or italics or color differences or any other possible differentiators.
Anyone dissatisfied with this, either needs to supply his own, better suited callable or can choose to deactivate that mechanism altogether by setting parameter hdr_info=False.

[JorjMcKie]

I have made a few more changes:

As per header identification, I have introduced a new parameter that limits the amount of accepted header levels to something smaller than 6.
When executing the following modified script:

import pymupdf, pymupdf4llm, pathlib

filename = "test.pdf"
doc = pymupdf.open(filename)

# pre-process header identification and only accept 4 header levels.
hdr_info = pymupdf4llm.IdentifyHeaders(doc, max_levels=4)

md = pymupdf4llm.to_markdown(
    doc,
    # write_images=True,
    force_text=True,  # irrelevant because of below
    hdr_info=hdr_info,  # use the prepared header info object
    show_progress=True,  # progress bar
    ignore_images=True,  # ignore all images
    ignore_graphics=True,  # ignore all drawings
)
pathlib.Path(doc.name).with_suffix(".md").write_text(md)

You would need the wheel
pymupdf4llm-0.0.20-py3-none-any.zip

And this is the result:
test.md.zip