DRAFT: Feat/contact details

[metalwarrior665]

@foxt451 Creating a draft PR for better review

[jancurn]

Great stuff, I'm curious to see how it will work. Surely the LLM will have some advantage over regexps, no?

[metalwarrior665]

I don't think so, all social handles have pretty clear regexes. Emails also should be fine. Phone numbers can be tricky but I don't expect GPT to do a better job than regex, it will also halucinate more. GPT is for things that require an understanding of a broader context. Let's continue in the Issue

[foxt451]

Yeah, I realised a bit later that this offers no real advantage over regexes. I guess we could set up some metamorph chain between contact-details scraper and gpt so that gpt gets page's text, regex-extracted contacts and just matches them

[metalwarrior665]

I would try the name + contacts grouping. That was the student project. The tricky part was actually finding enough testing websites because it is not as common.

[jancurn]

Yes, for extracting email, URLs, etc. it has no benefit, but the whole point of using LLMs was to:

• cluster the found handles to a specific people, and ideally also determine their full name
• be able to correctly identify phone numbers, ideally filling the country prefix - regexes are pretty much useless for this