Repository navigation
PDF to Markdown doesn't preserve text relationship or indentation #83
Description
Here's a sample PDF - https://www.bnm.gov.my/documents/20124/963937/Risk+Management+in+Technology+(RMiT).pdf/810b088e-6f4f-aa35-b603-1208ace33619?t=1592866162078
However there are several parsing errors which I will try to highlight below
Line elements aren't preserved
Output
G 8.5
To promote effective technology discussions at the board level, the
composition of the board and the designated board-level committee should
include at least a member with technology experience and competencies.
Indentations are ignored and text ordering is altered
The TRMF must include the following:
(a) clear definition of technology risk;
(b) clear responsibilities assigned for the management of technology risk at
different levels and across functions, with appropriate governance and
reporting arrangements;
the identification of technology risks to which the financial institution is
exposed, including risks from the adoption of new or emerging
technology;
(c)
(d) risk classification of all information assets/systems based on its criticality;
(e) risk measurement and assessment approaches and methodologies;
(f)
(g) continuous monitoring to timely detect and address any material risks.
List elements are ignored and presented in entirely new lines
1.
2.
3.
4.
5.
6.
The assurance shall be conducted by an independent external service provider
(ESP) engaged by the financial institution.
The independent ESP must understand the proposed services, the data flows,
system architecture, connectivity as well as its dependencies.
The independent ESP shall review the comprehensiveness of the risk
assessment performed by the financial institution and validate the adequacy of
the control measures implemented or to be implemented.
The Risk Assessment Report (as per Part D in Appendix 7) shall state among
others, the scope of review, risk assessment methodology, summary of findings
and remedial actions (if any).
PDF parsing will always be a PITA, but I think these issues can be addressed by tracking the locations of the elements, right now I feel it simply loops over the textual elements and uses simple algorithms to merge them together
Activity
[-]Markdown doesn't preserve text relationship[/-][+]Markdown doesn't preserve text relationship or indentation[/+][-]Markdown doesn't preserve text relationship or indentation[/-][+]PDF to Markdown doesn't preserve text relationship or indentation[/+]Utsav-Mehta commentedon Feb 17, 2025
Is anyone working on this?
Utsav-Mehta commentedon Feb 19, 2025
Gagan Bansal (@gagb) is anyone working on this? Feel free to assign this to me!
gagb commentedon Feb 19, 2025
Assigned! Thanks for taking a look! We are a small team and this is an OSS project, so contributions are very welcome. Thanks again!
Utsav-Mehta commentedon Feb 19, 2025
Thanks for assigning Gagan Bansal (@gagb) . It's a pleasure, I am considering integrating vision models to handle relationships and indentation. Let me know if this sounds like a good starting point and if you have any suggestions!
gagb commentedon Feb 19, 2025
Sounds like a good experiment. Creating a plugin would be a good first step. Recommend using open source vision models.
Utsav-Mehta commentedon Feb 20, 2025
Thanks for the insights, will be creating a plugin first as you mentioned.
iamskp11 commentedon Sep 16, 2025
Hey Utsav Mehta (@Utsav-Mehta) , if you are not actively working on it, I would like to pick this issue. Gagan Bansal (@gagb) You can assign if possible.
cagdasyurekli commentedon Aug 26, 2026
I opened PR #2345 with an opt-in semantic PDF OCR mode aimed at the hierarchy, indentation, list, and reading-order failures described here and in #41.
It is isolated in the existing
markitdown-ocrplugin and disabled by default. The prompt requests faithful Markdown structure while treating document instructions as untrusted data; failures or empty page results conservatively fall back to the existing deterministic whole-document conversion.Tests cover nested-list/table Markdown preservation, constructor and per-call overrides, both PDF rendering backends, fallback behavior, resource cleanup, and compatibility with legacy one-argument OCR services. Live-model output remains model-dependent, so this is intentionally an explicit alternative rather than a claim of perfect recovery for every PDF.