Add source infos: text chunk and its page number, possibility to hide document link

[PsychicBirdy]

Hi, congrats to such a well-conceived app. I have a feature request. Currently when using sources, h2oGPT adds to its reply a score and a link to each document used. What I would like to see added here:

• A snippet of the actual document chunk used to create the reply. GPT4all and localGPT both offer this. In GPT4all you click on a "context" link and the text chunk pops up. It's not copyable though. It would be an improvement if the chunk could be even made copyable.
• The page number the chunk can be found on. Again GPT4all already shows this, see the screenshot.
• A way to deactivate the link to the document that h2oGPT currently provides (but keeping the document's name, page number on its own makes no sense, of course!). Why is this a desirable feature? Sometimes laws only allow to data mine documents and use them in AIs but prohibit to share the original documents in full. That is, if the GPT app is provided publicly, for example. That's why it would be great to be able to deactivate linking to documents but still have a way to inspect the sources.

[pseudotensor]

Sounds like a great idea. Right now I only provide link to source document, but a link to chunks and meta data would be useful I agree.

[PsychicBirdy]

Thanks a lot for looking at it! Being able to access those snippets would also make replies much more transparent and comprehensible, the one thing basic chatbot replies are currently missing.

[pseudotensor]

It's maximally hidden in accordion, and has the actual snippets now. I'll add other metadata soon.

[PsychicBirdy]

I tested it and it looks good. Thanks a lot for implementing this! The accordion is a good idea so it doesn't get too cluttered. Now I'm only missing the page number and a switch to deactivate the linking to the document.

[pseudotensor]

FYI @PsychicBirdy you can see full meta data from database here, although not in references yet:

[pseudotensor]

@PsychicBirdy

I added the switch to show url or not. Before:

with --show_link_in_sources=False

[PsychicBirdy]

Looks good. However, I thought of the switch in a different way: When set to False, I would keep the path and filename - without them the other source information is kind of vague - and only remove the linking to the files. So for example "user_path_many/HAI_AI-Index_Report2023(2).pdf" is still shown, but you cannot click on it anymore, disabling the possibility to download files.