Parallel Picture Descriptions via API

[vaclcer]

Hi, I am using PictureDescriptionApiOptions to describe images in input documents via external vision model (running in vLLM).

I found out that descriptions generation takes most of the time of processing. My question is: is Docling calling external service as it stumbles upon an image to describe it? Is this call blocking or not?

It would be a smart move to go through the document first, gather images and then describe them in parallel, maybe in batches.

If it works like that already, is it somehow configurable? Thanks!

[dolfim-ibm]

This is indeed already possible with the concurrency argument of the PictureDescriptionApiOptions object: https://docling-project.github.io/docling/reference/pipeline_options/#docling.datamodel.pipeline_options.PictureDescriptionApiOptions

[vaclcer]

Awesome, thanks a lot for a swift response! If this works as I think it should, that would make a massive speed gain and default

concurrency: int = 1

seems very conservative!