Activity table recognition inaccurate problem

[jelymewang]

If there are two tables on one page, splitting the table can improve the accuracy of md. Is there a good way to split the table?
At the same time, for the case where there are no activity names but only numbers and values starting from the second page of the activity table, the large model is inaccurate when there are multiple values for one number. Can it be solved by merging md or context mechanism?

[jelymewang]

The inaccurate situation of pdf table to md has been improved by Mistral OCR. The extracted table structure is clear and error-free. The only disadvantage is that it costs 1,000 pages/$.
The inaccurate extraction of multiple active names corresponding to one compound number by the large model has been improved by modifying the prompt and replacing a better language model.

[dahuilangda]

Thanks a lot for your valuable insights and updates!

If you're open to it, feel free to submit a PR — we'd really appreciate your contribution!

[dahuilangda]

Thanks for the great suggestions, @jelymewang!

Both dots.ocr and the Mistral OCR API are excellent ideas. Based on your feedback, I plan to integrate both as options in the next version.

I really appreciate you sharing this. Stay tuned for the update!