search

msull / emilytarot

AI Tarot Readings powered by Streamlit and GPT
https://emilytarot.com
MIT License
23 stars 5 forks source link

Code does not continue after reading, with further query #2

Open forrie opened 7 months ago

forrie commented 7 months ago

I'm running the latest revision here, and the code will end with a question from "Emily" but it summarily closes down the reading/chat so you cannot continue. This seems like a bug.

msull commented 7 months ago

Hi there -- are you running your own local copy of this, or accessing emilytarot.com?

How did you come across this repo? I don't think I've advertised it anywhere but on my linkedin profile, its was just an experiment I did to learn streamlit and LLMs last year, though I've certainly thought about expanding Emily and particularly supporting alternative models like Claude to see how it alters the readings.

forrie commented 7 months ago

Thanks for replying.

I am now running my own private instance, under Docker in Linux (via MacOS), using my own OpenAI key. I'm enjoying it, and I see a lot of future with this. I was discussing something like this with someone else a while back, as I have been submitting card draws to ChatGPT -- without specific format instruction, it does an OK job interpreting.

How did I find the repo? I think it was via Google. I can't recall.

I'd love to see longer readings like Celtic Cross. It would also be easy for the end user to provide his/her own imagery (for tarot cards), perhaps add the visuals for the draw, etc. Cosmetics.

I'd also be interested in how the other engines interpret. I did at one point try comparing ChatGPT and Bard (no Gemini) and they seemed pretty similar at the time, in their interpretation. I've made sure to draw the cards separately, as I don't really trust that the AI is truly objective -- they could be given instructions to always keep something positive, etc. You know what I mean.

forrie commented 7 months ago

I encountered another error where the chat session halted without continuance -- it was about to explain the draw of three cards. It started and then ended with "link to session" without any further continuance or prompt. FYI. It's not yet clear to me where the bug is -- it may be something in the API calls?

msull commented 3 months ago

I've recently updated this repo to use the new GPT 4 Omni Mini model -- this model is significantly smarter and better at following instructions, I'd highly recommend you update as well.

I saw several sessions that terminated as you have indicated, and in all cases it wasn't necessarily a bug just GPT 3.5 not sending any further actual user questions or tarot instructions in the response, which is how the app recognizes that the session has ended. I'd wager that 4 Omni Mini is much less likely to send a response like this -- very incredible model.