a16z-infra / ai-getting-started

A Javascript AI getting started stack for weekend projects, including image/text models, vector stores, auth, and deployment configs
https://ai-getting-started.com/
MIT License
4k stars 643 forks source link

Text search doesn't seem to work #30

Closed daaain closed 1 year ago

daaain commented 1 year ago

Not sure which part isn't working, but it can't pick up a question that is in the article word by word.

Screenshot 2023-06-23 at 11 54 57 Screenshot 2023-06-23 at 11 55 05
timqian commented 1 year ago

Meet the same issue, maybe the data on supabase is somehow lost? @ykhli

Screenshot 2023-06-24 at 15 12 56 Screenshot 2023-06-24 at 15 13 03
ykhli commented 1 year ago

Ah you are right. I just re populated that table. try again to see if it works now?

timqian commented 1 year ago

Works now!

ykhli commented 1 year ago

@timqian I think @daaain was referring to their local environment and not the demo instance. Since the answer to the question should not have been "I don't know", as the blogpost mentions "AI will not destroy the world..."

I suspect this is either because of the vector search results weren't optimal or because of chunking problems. @timqian in case you have time, feel free to look into it - otherwise i can look into this tomorrow!

daaain commented 1 year ago

I was looking at the demo instance, but if you could write up a few troubleshooting steps why this might be happening locally and how to fix it that would be really appreciated as I also would like to know how to run it! 😸

gsmckenzie commented 1 year ago

I had this issue with my local version, but it was because I had run out of free OpenAI credits. It is worth checking the OpenAI usage (https://platform.openai.com/account/usage) to check that you have credits. No credits will also cause the embeddings to time out with a retries error.

Obviously you issue may be different, but creating a new OpenAI account (with a different phone number) solved this for me.

ykhli commented 1 year ago

Great @daaain I actually tried this locally and didn't reproduce the error. Also confirming this was fixed on the demo instance.

Thank you @gsmckenzie @daaain!!