OpenAI Assistants in Bubble - A first look 

Thanks for the kind words!

Glad it was helpful!

Glad it was helpful!

@@LaunchableAI I've been experimenting with RAG quite a bit (Bubble/Pinecone/Flowise) - I see these assistants only allow for a max of 20 files to be attached to a given assistant and I believe each file has a max of 100k characters. My initial thinking is that the implementation of RAG inside an assistant isn't ideal in that there does not appear to be any method of controlling or directing the retrieval process (as compared to Pinecone/metadata for instance). I'm keen to know what your thoughts are on this. I'm tending toward experimenting with creating a tool that an assistant can use where it 'hands off' the user queries to a 'Pinecone tool' along with a prompt explaining the tools role in the whole RAG precess (it simply returns relevant chunks and document references) which the assistant then uses to synthesise the response as per a typical RAG process.

​@@sitedev You make some excellent points. I was talking to a client about exactly this this morning. We've built a bunch of pinecone-based storage & pre-processing bits, and are thinking about the Assistants + Files APIs as replacements. My thoughts currently are in line with yours. There's not quite enough flexibility with the current OpenAI options, for some more complex use cases (e.g., we're doing database and S3 retrievals with LangChain, passing to Pinecone; this sort of thing isn't an option yet. I can imagine other cases too that wouldn't work). That being said, I suspect they'll increase the file limit and the per-file size limit over time, so perhaps its not the right option for some projects now, but will be more viable soon enough. There's also some concern about vendor lock-in, if you do all your data storage and indexing via OpenAI, it makes it tougher to use other models / platforms. So depending on your industry / use case, something else to keep in mind. And lastly, you're point about passing some helper context to the assistants API, and using pinecone as a "tool", is probably an excellent idea. I hadn't though of turning pinecone etc. into a tool that the ChatGPT could call directly, but it sounds like a topic that's ripe for a tutorial video ;) If you try it, please let us know how it goes!

Glad it was helpful!

Yeah I ran into this problem too and spent awhile figuring it out. Thought initially the call had to be made client-side, so tried that. Later on in the tutorial (might be part 2), I think I discuss that you can't actually make the GET threads request; it's not supported. You need to store your thread IDs on your own, and use those. Last I looked, there was an open issue on OpenAI's developer/community forum of people discussing (ahem complaning about) this.

@@LaunchableAI The same issue still exists :( Just gonna continue with the video and hope you say how to fix it lol

I think I fixed the issue. Simply change it from GET to POST.

Yep, the latest plugin versions and recent tutorial videos cover streaming. May also be releasing a tutorial on how to build streaming from scratch

If there's interest in expanding, maybe I'll make some more on the topic, sure.

You can try using a plugin. Our plugin "ChatGPT Toolkit" has various functions for extracting text from files and websites. Maybe that would do the trick? It's a paid plugin ($10/mo), but you may be able to find some free alternatives if that price is too high. If you don't want to use a plugin, you'll probably want to find an API that can accept files or websites and return text. Then you''d pass that content to ChatGPT when you make a request.

Yep, it's kind of pricey, esp. if you're outside North America or Europe. I use gpt-4 pretty much every day for my work, so it's worth it for me, but I can see that it wouldn't be worth it if you're only using it occasionally or only for this one feature.