Is it worth to upload embeddings as knoweledge?

I’m on preparing data for upload as knowledge to customGPT.

  1. Do you mean, is it worth, to calculate embeddings and upload them together with knowledge textual content?
  2. If yes, how would you format the upload file? My guess is a csv, with the textual content in A, and corresponding embeddings in B.

I’m only guessing, but I think uploading embeddings with your CustomGPT knowledge files would be redundant in the format you suggest.

Again, I’m only guessing, but I think the CustomGPTs have their own limited Vector Store for file storage, since they’re basically limited versions of Assistants. Vector Stores automatically add embeddings and keyword fields.

That said, I’ve been wondering if a meta field of embeddings of keywords and concepts, either added to a Knowledge File as meta or as a section of text at the top of the file, wouldn’t help create or reinforce the desired web.

1 Like

I don’t think that makes sense because GPTs do not have themselves the possibility to compute embeddings for search terms unless you provide this function through a custom action. So, all the GPT could do would be a similarity search between documents, not a semantic search related to a query.

@thinktank Really, have you ever seen this? Afaik accesses of GPTs to files are transparent through use of the code interpreter? I have never seen it uses any semantic functionality.

Nope. I’m just guessing based on the similarities between the two systems.