When one wants to speed read scientific articles or books or listen to it and one doesn't want to use a LLM that was trained on transforming it directly, one has to do it in steps. For the first task, this is necessary anyway and, furthermore, one can do some rearangement like I was tought to “speed read” an article—basically reading important parts first to avoid reading it at all at an early stage or if one reads it to remember the more important parts better.
I once read an article about Docling by the famous Simon Willison and tested it.
More recentyl I stumbled upon an article comparing Docling, Marker and LlamaParse by Khuyen Tran which suggested Marker to me for my purpose.
So let's test it.
I tried a paper with a quite complicated first page. Indeed, that page wasn't rendered well using Docling, but ok with Marker. Docling failed with the metadata table and merged its values into the text. However, Marker sorted page per page and doesn't include the figure captions, so that the images are at seemingly random posissions and can't be matched well.
When it comes to the code rendering, Docling at least recognises that it should be rendered as code, though, Markling gets the content a bit better. But neither version is good enough to be used without manual transcription/intervention.
So, I might have to read a bit of their documentation to figure out if I can use the structure of one of them and the OCR and tables from the other. Or I'll have to find a different solution, maybe testing the local AI models which I tested also against a similar, but way easier case.
Another promissing software, I haven't tested yet, is xberg the successor of Kreuzberg.
Websites
For websites, there is a commercial service with a free variant called Jina Reader which claims to transforms Websites of any kind to LLM friendly text. It also reads imgaes and PDFs. But similar to the other two tools, it doesn't do what I want it to when providing it the difficult PDF. It ignores code blocks, preserves the site splitting, forgets about formating headings.
Related
Somewhat related, are systems that index your documents and search on them like Perplexity does for the web or some local LLM provdiers like litellm do. Such a system is localGPT. So far, I haven't tested it yet, so I can't judge it. Also I haven't had a look into its source code to figure out how they do the parts I am interested in, but they also seem to use Docling.
Between localGPT and Reader is unstructured. This commercial service (free tier available) claims to structure ones data of any kind. Their focus is on many different (50+) formats (docs).