partition_pdf should support documents in two column format

[MthwRobinson]

The goal of this issue is to add support for two column documents to partition_pdf. Currently, partition_pdf will process the documents but the elements are not always in the correct order. The attached document can be used for testing.
Dense-Passage-Retrieval-for-Open-Domain-Question-Answering.pdf

[hsm207]

does this issue also implies adding support for docs with more than 2 columns?

[MthwRobinson]

@hsm207 - Yes, this will support multicolumn (two or more columns). In the meantime, you can pass in strategy="fast" (if the text is extractable) or stragey="ocr_only" (if the text is not extractable) for better results on multi-column documents.

[CShorten]

This will be huge, so excited for what this will unlock with putting scientific papers into LLM systems!

[Kate-Lyndegaard]

Hi,
Is this improvement available, yet?

[walsha2]

Is this improvement available, yet?

In tracking the commit history in the unstructured-inference repo, it looks like a promising first attempt was added in #112 and reverted in #120 - then no action since then I presume?

Maybe @MthwRobinson or @qued can comment on the status of this. Is a more sophisticated sorting method the right approach or is that too fragile? Seems like sorting is required no matter what, esp if the text is directly extracted from a pdf.

@MthwRobinson was on the right track and I think a few more cracks at "smart sorting" might yield a solution that is at least do no harm to single column layouts while also addressing some larger set of multi-column PDFs.

Not sure if there is some other means of solving the multi-column problem that is being tossed around internally but has yet to be disseminated? Wanted to ask first before digging any deeper myself.

[cragwolfe]

Yes, there was a significant improvement in reflecting "natural reading order" in #1161 , which was released in 0.10.6.

However, there are a couple of open issues to further improve on this:
#1233
and
#1209