Prodigy-PDF tutorial: PDF annotation and OCR correction

This tutorial, uploaded on October 24, 2023, demonstrates two Prodigy-PDF recipes for annotating documents and correcting OCR output. The source does not specify a software version or a separate recording date, so the workflow reflects the demonstrated 2023 setup.

The speaker starts with an academic article about StarCraft algorithms. The PDF image manual recipe converts PDF pages into images for region annotation and processes PDFs from a selected folder. The example labels titles, paragraphs, and figures, then saves the annotations to a Prodigy dataset. Zoom controls help the annotator inspect boundaries around each region.

The PDF OCR correct recipe reuses that dataset to extract text from selected regions. The speaker chooses titles and paragraphs, leaving figures out of this pass. The correction interface displays the region label, its cropped image, and the OCR result, which the annotator can accept or edit.

The speaker attributes the example's good OCR results partly to high-contrast, black-and-white pages and internal upsampling of the selected regions. The output still preserves dashes at line endings. A setting called fold dashes merges the text into one string and removes those word breaks in the demonstration. Whether that cleanup suits a document depends on the intended use; the example does not establish OCR accuracy for other PDFs.