Optimizations to the Docling Codebase

[agodinezmm2007]

Hello all,

Thank you for open sourcing this library. Docling is a central component of a project I have been working on. I have used docling to extract textual data from PDF articles for integration into graph databases, considering i was previously using Azure Document Intelligence for this the cost reduced dramatically to $0. There were various bottlenecks however, which necessitated multiple modifications and optimizations to ensure multi GPU usage and increase throughput. I have created a detailed report of the integration of Docling into my project, the strengths of docling, its weaknesses, and how I addressed or attempted to rectify those weaknesses. I would like to do pull requests for all of the files I have modified, however many of the files were heavily refactored. Please visit the following section for my analysis of docling:

https://agodinezmm2007.github.io/project_portfolio/05-technical-report.html#stage-4-content-extraction-via-document-layout-analysis

This video is sped up 8X but at 14:20 you can observe doclings' performance https://youtu.be/ZCy5ESJ1gVE?si=ADU436JwfM5pvLod&t=862

Docling processed 100 academic journal PDF articles in 3 minutes, which includes text, tables, and formula extraction. One of the reasons for the speed bump is because instead of waiting for SmolDocling to reach its token limit when stuck regenerating/spamming characters I included logic within code_formula_predictor.py to detect these loops and stop them. Of course, this results in missing some formulas but for now its as good I could get it to perform.

[DLME2024]

Are you interested in doing any freelance work for a project I am working on? We wouldn't be analyzing academic papers but rather mortgage loan documents, pictures, pay history, etc. Essentially a document refinery. Put in virtually any type of content and recieved beautifully structured data/database records created/flattened & cleaned pdf's. Organized, separated, collated, etc.

[agodinezmm2007]

Sounds interesting, is it more text extraction or context extraction/classification?

[DLME2024]

It’s both. Equally important in my use case. There’s also a refinement piece to it as well where we are splitting bulk docs, orienting/collating, finding page ranges, and finally created nice clean PDF’s as output regardless of file input type (aside from pictures).Dominic McFadinOn Aug 1, 2025, at 12:11 AM, agodinezmm2007 ***@***.***> wrote: Sounds interesting, is it more text extraction or context extraction/classification? —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you commented.Message ID: ***@***.***>

[agodinezmm2007]

It’s both. Equally important in my use case. There’s also a refinement piece to it as well where we are splitting bulk docs, orienting/collating, finding page ranges, and finally created nice clean PDF’s as output regardless of file input type (aside from pictures).Dominic McFadinOn Aug 1, 2025, at 12:11 AM, agodinezmm2007 @.> wrote: Sounds interesting, is it more text extraction or context extraction/classification? —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you commented.Message ID: @.>

Email me at [email protected] with project details. Send 30+ PDF samples (5 per format if diverse). Clone https://github.com/agodinezmm2007/docling_mod and test with scripts/sample.ipynb to see if it handles your document types effectively.