CacheChat

[snow garnet]

Hey everyone, just finished a short mid-week hack for a website that allows you to talk to GPT-3.5 about one or more URLs and files (of nearly any format and size!): https://cachechat.pagekite.me/. The best part: everything's built simply with numpy and OpenAI's API, no vector DBs or anything overcomplicated. If you like what you see, please star the repo: https://github.com/andrewhinh/CacheChat

[snow garnet]

Thanks for pointing it out. It’s a currently known issue, and will be fixed soon

[snow garnet]

Alright website fixed, thanks

[urban harness]

Can it work with contents of websites, not just articles URLs?

Because my output looked generic instead of content-specific to the URL

[snow garnet]

Yeah that seems to be the work of a minimum similarity requirement between the prompt and sources, guess I should remove it for the sake of general questions like this. Again, thanks for pointing this out!

[snow garnet]

Issue fixed, thanks!

[urban harness]

@snow garnet your app didn't parse a URL I inputted ||https://www.kbp.org.ph/accreditation/|| for the question.

The prompt was "what requirements are needed"

[snow garnet]

Let me check it out

Well that's an invalid URL:

Though yes it should've failed earlier

[urban harness]

Thank you for pointing it out. It's a server issue 😦

I tried it on another one -- https://www.liquor.com/brands/tequila-rose/

[snow garnet]

How'd it go?

[snow garnet]

Oh interesting, will look into it, thanks

[snow garnet]

Btw, it's hard to fail early for a URL since web parsers get easily confused even with bad web pages. Will look into how to make this better

[snow garnet]

Ah, seems I forgot to do a unit test case. Fixing it now

[urban harness]

Literally ah

[snow garnet]

Turns out that website had a TON of junk, which made the context window overflow for the first question. Those BeautifulSoup web parsers are nice, but they get confused sometimes 🙈

[urban harness]

Yes, sometimes some engineering is needed to get the gist

[snow garnet]

Yup, fixed now

[urban harness]

Tried also a jpg file with text and a docx file. But there is another error

[snow garnet]

Ah yes the issue of no text being found, on it

[snow garnet]

Fixed it

[urban harness]

Headache

[snow garnet]

Nvm, now it's fixed 😎 Apologies for the confusion 🙈 Had a logic error that wasn't complicated matrix math!

Thanks for the tests!

[urban harness]

You could have provided acceptable file formats as a list so people would be oriented about which to upload.

[urban harness]

My pleasure.

[snow garnet]

Yep you’re right, thanks

[fair cobalt]

Really love this idea but having trouble ingesting documents when running my own instance. Your hosted website had no problem ingesting this PDF and other documents but I couldn't upload after running streamlit

[snow garnet]

Thanks for trying it out! That’s weird since I have the current main branch code deployed to the website. Mind sharing more of your error logs?

[urban harness]

@snow garnet Does Beautiful Soup parse Scribd documents? I have a Scribd document link but it shows a paywall. Trying Cache Chat for the document it won't access the part I was looking for

[snow garnet]

Yeah any url that isn't public, free, and valid won't work

[urban harness]

https://www.scribd.com/document/450192656/10-poems-of-Rio-Alma-docx

A workaround for Scribd documents should exist. I found a downloader but it already broke.

[snow garnet]

The problem is they state To access this document, upload one of yours, or subscribe with a 30 day free trial. Trying to get around this security measure doesn't seem right to me.

Of course I could implement logging in with Google, etc. to get this to work, but that isn't really the point of the project

However, I do feel a bit inspired to build a production-grade version of this tool.

[urban harness]

Login-to-use-this-project shouldn't be necessary

[fair cobalt]

Ahhhh my log would only output the "package punkt is already up-to-date!" message and it looks like I get that same output even when file ingestion is successful. I am a quite 'code illiterate PM' so turning on better Steamlit logging proved.....difficult for my small brain. Problem is certainly PEBKAC here though because I re-did the d/l and step-up on a different machine than my primary and it worked just fine

Also, again, code illiterate product manager here - I couldn't get "export PYTHONPATH=." to work on my windows system so I replaced it with "set PYTHONPATH=.", but, couldn't figure out how to modify "echo "set PYTHONPATH=.:$PYTHONPATH" >> ~/.bashrc" to work for me since my system couldn't find a ".bashrc" file, so I wouldn't have to do set the path each time I am starting this up - NOT expecting you to solve this for me but just sharing so you know how dumb some users are 😃

[snow garnet]

No that’s a fair point, I should make it inclusive for windows users. Thanks for bringing it up.

[snow garnet]

Actually, does replacing set with setx make the path change persistent? If so, happy to add this to the repo README

[fair cobalt]

I'll check

[fair cobalt]

So I couldn't just use "setx PYTHONPATH=." but tried to use "setx machine PYTHONPATH=." and "setx mypath PYTHONPATH=.", however, neither seemed to make that path persist once starting up another conda env.

[snow garnet]

Yeah that’s the issue with Windows, it’s never consistent even between other windows machines. I’ll add a link to a website I found for info around this in the README.

Comment can be found in step 2 of the setup

[fair cobalt]

Thank you for the help friend

[snow garnet]

Not sure, though it seems to still be able to give an answer

[urban harness]

@snow garnet Did your page got removed?