docling-project / docling-project/docling
Discussion: classification-based filters for picture description
- Dominant language
- Python
- Stars
- 66.4k
- Forks
- 4.8k
- Avg merge
- 2d 21h
- Merged PRs (30d)
- 84
Description
Hi all — opening a short discussion to explain the motivation behind PR #2836 and get feedback.
In our pipeline we classify pictures (e.g. chart/table/logo/screenshot). We only want to run picture description on a subset of those classes, and optionally enforce a minimum classification confidence.
PR #2836 adds three optional fields to PictureDescriptionBaseOptions:
- classification_allow
- classification_deny
- classification_min_confidence
and uses them to skip describing images that don’t pass the filter. This is especially helpful to avoid expensive descriptions for e.g. logos, screenshots, or low-confidence classifications.
PR: https://github.com/docling-project/docling/pull/2836
Would love feedback on naming, behavior, and whether this belongs in core options. Happy to update docs/tests as needed.
Contributor guide
Assessment
This issue has not been assessed yet.