baidu

/

Unlimited-OCR

like 3.9k

Follow

BAIDU 3.65k

Hi,

I'm using Unlimited-OCR for document parsing and have noticed that the generated Markdown output is already post-processed. It appears that functions such as remove_det() (or similar post-processing) strip the original detection tags before the results are written to the output file.

Currently, the output is plain Markdown/text, for example:

Introduction

This is the first paragraph...

However, I need access to the original structural predictions from the model, such as:

<|det|>title<|/det|>Introduction
<|det|>text<|/det|>This is the first paragraph...
<|det|>table<|/det|>| Col1 | Col2 |

Having these tags would allow me to distinguish titles, headings, body text, tables, captions, formulas, and other document elements so I can perform my own custom post-processing and generate Markdown according to my application's requirements.

Would it be possible to add an option such as:

model.infer_multi(
    ...,
    save_raw_output=True,      # Save output before remove_det()/post-processing
)

or

model.infer_multi(
    ...,
    post_process=False
)

This could save the raw decoder output (including the <|det|> tags) alongside the existing processed Markdown output.

This feature would be very useful for downstream applications that require document structure information instead of only flattened text.

Thank you!