Hi
I’m building an app / website for a sports (paragliding) club, part of which is building a knowledge base of interesting articles and links, which I’ve done - so far so good.
I’ve connected this knowledge list to ChatGPT which gives me a rather splendid way to search the database; screenshots below if anyone is looking to do this too. I’ve utilised the ‘answer questions about a table’ feature which gives me the sort of response I’m looking for i.e. one fairly tightly confined to answering questions about the table.
The club uses Telegram extensively for communication and discussion and I’d like to be able to connect this to my knowledge base; probably through make although I’m open to suggestions.
Currently I’ve been able to connect make to Telegram,
I can send messages from inside Glide to Telegram:
Before doing this you need to set up the integration with your OpenAI API. You don’t need a premium account but do need tokens - for testing a few £ did it for me
The context from the chatbot table is very important; this contains a set of fixed information about how I want the response to be generated; things like present it in markdown, always give a link to the relevent table item
Hope that helps someone. As I said above I’m trying to replicated this functionality with the question submitted from telegram and returned to telegram, probably using make.com if anyone can help with that
With “answer questions about a table”, what Glide does under the hood, as far as I can recall, is first converting your table into vectors, then your question to a vector, and pulls out the most related vector from the table to try to answer your question.
Now I don’t have actual experience with building something like this, I did try using OpenAI’s embeddings last year but it was more of a personal project, and using Python, not via their API.
This Vector Store thing, going alongside their Assistants API, looks like what you need to build an application for that.
However, Glide is handling the “ever-growing database” part for you in the app, in the Assistants API you’re uploading files for them to convert to vectors, so how you handle adding new records to the vector store, I imagine that would be a headache.
@ThinhDinh didn’t David just say in the annual update that this year they would look at integrating RAG into Glide Tables? Or did I misunderstand that?
I have a few use cases for RAG and have built a pinecone integration via make.com myself and another one directly from glide with stack-ai, however the pinecone one is (imo) a bit more complex to setup and maintain and stack-ai doesn’t give me the granularity for settings and metadata filtering to get great results out of it.
If you or someone would have more info about Glide integrating RAG I would love to be involved!
Sorry, I’m out of the loop on this, can you share where he said that? If it’s a meeting, 100% I missed it because they’re usually midnight to 3AM my time.
Also, my comment was a bit old. Since then, internally at my agency, we have tried various third party solutions that can accept links from our knowledge base and answer our new hires’ questions. I think that’s the closest use case I have for this.
I’m just a fellow user like you, I almost have no inside info
No problem, I assumed you would know more. You seem to be involved in everything ;-), probably due to all the great advice on the forum here.
But anyways I’m looking for better tooling to get RAG quality at a decent level.
Just random chunking (without adding context, metadata filtering & reranking) and storing in a vector database doesn’t seem to be enough. But I didn’t find a nocode tool yet that allows you to do that via an API.
Thanks for the link. That does sound like what you’re suggesting as well.
However, I believe that isn’t necessarily a new thing, the “Answer question about a table” action already does that. I believe they just want to enhance it.
Maybe but if you rewind to 10 seconds before he says it’s an integral part of their underlying database infrastructure. Also for big tables. So I’m hoping/assuming it’s a big thing.
