pdftext

There were some cases where visual and text coordinates didn't align. This fixes that issue.

• There was a bug where pdf paths were assumed to be strings - this is not always the case

• Faster inference with ONNX
• Remove warning when loading scikit-learn model
• Flatten form fields into pdf

Fixed bug that didn't unnormalize bboxes properly.

• Optimize dictionary access and loops to get an ~10% speedup

Enable optional parallel workers when extracting text. This can cause a performance hit on small pdfs, but can speed things up 2x or more on larger ones. This can be done with the --workers flag via CLI, or via the workers kwarg.

Not all spans would have the right font information before. This fixes the issue.

• Charbox has zero width/height when loose=True with rotation

• Fix logic for pulling font names
• Increase sample frequency

• Use line breaks from pdfium

• Option to keep characters with JSON/dictionary output
• Fix some bugs when interfacing with the pdfium c api (thanks @mara004)

• Add probability threshold for block predictions

• Select a range of pages versus converting the whole doc
• Minor internal refactor to use docs versus paths

• Fix bug where hyphens didn't show up at the end of lines
• Improve wrapping for hyphens - join words across hyphens before newline (disable by passing keep_hyphens)
• Restructure output to avoid redundant info in json blob - keep track of text spans with similar font info instead of individual characters
• Update model to predict blocks more accurately

• Switch the character box to a loose box, to get the full character range

• Rotate bboxes if pdf is rotated