feat: Use DeepWiki LLM to answer users questions

[Zeglius]

Allows Deepwiki to scan the documentation and provide users with a ChatGPT-like chatbox to answer questions.

[Zeglius]

Though, Deepwiki is more centered towards developer-centered documentation, if any of you can find something better tailored to an average person, would be nice.

cc @castrojo @KyleGospo @wolfyreload

[wolfyreload]

@Zeglius I tested it with more than 10 questions, some easy, some harder, and it got them all right. No AI hallucinations. For example, it told me it had no clue about "Bazaar" since that's not in the documentation. Downsides? It's definitely built more for devs than everyday folk. Can't think of any alternatives off the top of my head though!

[Zeglius]

@Zeglius I tested it with more than 10 questions, some easy, some harder, and it got them all right. No AI hallucinations. For example, it told me it had no clue about "Bazaar" since that's not in the documentation. Downsides? It's definitely built more for devs than everyday folk. Can't think of any alternatives off the top of my head though!

Yeah, it doesn't make shit up, treats everything in the repo as the only source of truth, which is right what we need.

Though, the first thing that pops up when you enter the website is all the documentation structured by the LLM, which in this case adds visual noise and may disorient the user, thus less chances of them seeing the chatbox (which is the main functionality that I want from it).

So yeah I'm in need of more streamlined solutions, if possible

[castrojo]

Thank you for investigating this, this is fantastic! I'm training dosubot on our official docs and it answers on Bluefin's discussions site: ublue-os/bluefin#2287

This means we could just have the bot answer. We should test everything and give options, I notice users using generic LLMs and they are not as nice as they could be.

I working on prototyping a local-first LLM that is trained on all our docs and source, and then ignores the general internet. We'd keep the prompt in git so the community can tweak it, but that's a ways off yet.

[Zeglius]

I'm training dosubot on our official docs and it answers on Bluefin's discussions site

That would require users to have a GitHub account, which is a huge turnoff.

[nicknamenamenick]

I'm training dosubot on our official docs and it answers on Bluefin's discussions site

That would require users to have a GitHub account, which is a huge turnoff.

We already have a discussions section in this documentation repository and there is encouragement to report issues in the Bazzite repository too. Of course, we know a good chunk of the userbase would prefer not to sign up, so I understand that. If we don't use Dosubot, then can we convince Kyle to remove the "Discussions" section on this repo?

Anyways, I'm for this especially since it may be able to answer a little more specific questions. Of course I still fear the bullshit it may spew since LLMs would prefer to come up with some explanation than going 'no that's not possible or a bad idea' by design. We can also change the top "block announcement" of the website to also mention using the bot to search for specific questions.

[asen23]

it will be nice if there is something like dosubot but for discord

[Zeglius]

We already have a discussions section in this documentation repository and there is encouragement to report issues in the Bazzite repository too. Of course, we know a good chunk of the userbase would prefer not to sign up, so I understand that. If we don't use Dosubot, then can we convince Kyle to remove the "Discussions" section on this repo?

didn't know we had discussions enabled in this repo. Not against disabling it, if anything there is discord/discourse. No reason to have so many baskets if there aren't enough eggs.

[Zeglius]

We can also change the top "block announcement" of the website to also mention using the bot to search for specific questions.

Would argue that's pointless. Ppl is going to ask whatever to the LLM regardless and adds more visual noise unnecessarily.
Using an LLM by itself doesn't guarantee 100% accuracy, and if the user doesn't know already I will assume their reading comprehension is not enough to read the notice.

[nicknamenamenick]

I added the DeepWiki link to the "Community Resources" doc until something happens here... if it does.