PDF to MD Conversion with Docling v2.18 is Incomprehensible

[sallahbaksh]

Bug

It looks like there was a bug introduced when converting from PDF to Markdown using docling v2.18. The resulting md file is pretty inaccurate.

Steps to reproduce

Use DocumentConverter() to convert pdf to md (Stafford County - VA Zoning Ordinance.pdf). The resulting md file looks like this:

Docling version

Docling version: 2.18.0
Docling Core version: 2.17.1
Docling IBM Models version: 3.3.0
Docling Parse version: 3.2.0
Python: cpython-312 (3.12.8)
Platform: Windows-11-10.0.22631-SP0
...

Python version

Python 3.12.8

[cau-git]

@sallahbaksh For comparison, can you please try with the PyPdfium backend? That makes it easier to tell if there is a fundamental issue with the PDF or if docling-parse needs improvements.

doc_converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(backend=PyPdfiumDocumentBackend)
    }
)

[sallahbaksh]

Here's the result with the PyPdfium backend:

It looks better than before but still not accurate. Didn't have this issue with docling v2.15

[saadkhi]

I would like to resolve this issue. Could someone kindly guide me on which file I should review to address the issue? Thank you.

[PeterStaar-IBM]

@sallahbaksh I will look into this. We basically need to fix a docling-parse. We will follow up here: docling-project/docling-parse#99

[dhdaines]

Hi, I have also encountered this problem. For those who might run into it, it is specifically due to some change in docling-parse 3.2.0. Docling 2.18 is actually okay, but you need to use docling-parse 3.1.2, which means you can't have docling newer than 2.18.

[dhdaines]

This should be fixed (though there is probably a long-standing underlying bug which remains unsolved) in docling-project/docling-parse#105