Invalid Sequences (e.g., /89, /.notdef) in Extracted Text

[zvictor]

Bug

When converting a PDF to Markdown using docling, the output includes unexpected sequences such as /89, /81, /.notdef, and other numeric or font-related codes interspersed throughout the text. These sequences appear to be PDF internal operators or font references and are not part of the original document’s content. This corrupts the extracted text and makes the output unusable.

Steps to reproduce

1. Run the command:

   docling --from pdf --to md --image-export-mode placeholder https://venda-imoveis.caixa.gov.br/editais/matricula/SC/1444409002480.pdf  

2. Observe the generated Markdown output, which starts with the following invalid sequences:

/89

/81

/89

/88

/118

/87

/82

/79

/34

/.notdef

/.notdef

/83

/83

/82

/86

/85

/81

/4

Docling version

docling 2.17.0

Python version

Python 3.12.0

Additional notes

The issue likely stems from the PDF parser incorrectly handling font mappings, operator codes, or undefined glyphs (e.g., /.notdef). These artifacts suggest the parser is either failing to filter out non-text elements or misinterpreting font/encoding data during extraction.

Therefore, this might be related to DS4SD/docling-parse#92.