Find answers from the community

Updated 2 years ago

I have a question which I should ve

At a glance

I have a question which I should've asked a long time ago... As most of you know (Logan), I'm building a podcast generative ai company, you upload your audio file and it generates things like summaries, titles, twitter threads, etc etc.

My question is: What is the best architecture to set this up so that you can access these documents and generate any number of prompts on it? Should I use something like pinecone?

10 comments

LLogan M

So the concern here is just how to manage many indexes/users right?

Yea, pinecone or weaviate are probably the top picks. If you want to host the vector db youself, there's also qdrant or chroma

bbmax

but, I don't even use embeddings at all right now, is that silly? Just an GPTListIndex (which maybe does embeddings in the background?)

LLogan M

Oh lol thats true 😅

Nah that's not silly, it seems like your use-case doesn't really require semantic search hey? (The list index doesn't use embeddings unless you tell it too lol)

In that case, you could persist each index as needed to S3 or a google bucket. We have this fsspec integration that makes this really easy
https://gpt-index.readthedocs.io/en/latest/how_to/storage/save_load.html#using-a-remote-backend

bbmax

I've been persisting to MongoDB. A semantic search key = choose an document based on a key?

bbmax

Thanks as always, and congrats 🙂

LLogan M

*hey, not key. Inner canadian slipping out lol

bbmax

lmaooo

bbmax

i'm 20 minutes from windsor in Detroit so I get it :).

bbmax

What about ensuring a response is a specific length? is that all prompt magic?

LLogan M

Sadly, that is prompt magic 🙃

Maybe you'd find some value in this guide openai shared recently?
https://platform.openai.com/docs/guides/gpt-best-practices

Add a reply