Currently the colorspace handling only supports DeviceGray and DeviceRGB and the handling is simplistic only looping through the images in XObject and compressing all of those. If any image was never used in the contentstream it would still not be removed for example.
Also this means that inline images are not handled.

The handling should be made more generic and use the ContentStreamProc

I noticed that there is no information on what column space actually means when using pdf_data().

The only reference I found so far is that its meaning might be unclear: https://discuss.ropensci.org/t/pdftools-2-0-powerful-pdf-text-extraction-tools/1520/4

How to reproduce:

• search for "président" in the demo website here
• open the doc entitled: "PE_HE_Angel_Mermaid_010610.pdf" ([this one](https://datashare-demo.icij.org/#/d/luxleaks/5499b664f9da72804ba28e6ff51d894abac7becc86ee2ad3b0a78c6a190eec1c6cb1586abb2dfbb17088b30117

I am struggling to figure out how to use this library to read a pdf as text for the purpose of Natural Language Processing as an alternative to

using Taro
Taro.init()
meta, txtdata = Taro.extract(files[1]);

as shown in
https://github.com/aviks/nlp-workshop/blob/master/NLP-in-julia.ipynb

Or can I not use

A short version of the documentation is available straight from Github (README.rst) while a more exhaustive one is present in the docs folder and online on trafilatura.readthedocs.io

Several problems could arise:

• Non-idiomatic use of English (not quite fluent or natural)
• Unclear or inc

Currently, it appears there's no check for whether the file has actually changed before rerunning textract so it probably reruns even if the user has only updated the title.

@gasman and I were discussing adding file hashing to Wagtail Images/Documents for cache-busting but might help solve this issue too.