Unstructured-IO / unstructured

Open source libraries and APIs to build custom preprocessing pipelines for labeling, training, or production machine learning pipelines.
https://www.unstructured.io/
Apache License 2.0
7.44k stars 580 forks source link

Suggestion: include consolidated bounding box coordinates in chunk metadata when using "by_title" chunking strategy #3194

Open NikitaKemarskiy opened 3 weeks ago

NikitaKemarskiy commented 3 weeks ago

Problem Currently when "by_title" chunking strategy is used and coordinates = true parameter is set (in order to return coordinates of the PDF chunks), coordinates are not returned (because in this strategy separate chunks are joined under the hood, which may span multiple pages).

"by_title" strategy is really useful because "default" strategy often returns really small chunks (containing one word or a couple of words). Therefore, inability to use coordinates with "by_title" strategy blocks use cases which require coordinates of text blocks in PDF files.

Suggestion The suggestion is to return consolidated bounding box coordinates when "by_title" chunking strategy is used, returning a rectangle with extreme coordinates of the included chunks if multipage_sections = False parameter is passed (therefore chunks cannot span multiple pages and Unstructured.io API can calculate bounding box coordinates on the single page).

Additional context The issue was discussed here: https://github.com/Unstructured-IO/unstructured/issues/1698

NikitaKemarskiy commented 3 weeks ago

As @scanny suggested, some "by_page" strategy can also be added (as far as I understood it implies "by_title" + multipage_sections = False)