The docker file runs a script to make sure that docling pulls its required resources into the image, as proposed in the above issue.
The python worker sends both the full document text as markdown and as chunks, stored in the same json on the COS.
I tested the supported document formats of docling via the UI to the extent that I could. However, the parser of the bee framework seems to trip over the (rich) document contents when they become part of the conversation. The document contents are retrieved properly from the COS though.
Implements https://github.com/i-am-bee/bee-api/issues/49