Generate embedding for a collection using individual embeddings

Community
,
1

Hi,
Problem statement:-
I have a collection of different chats per customer where I have individual openai embeddings for each of the chats in the collection. I have to generate the embeddings for the whole collection with the help of individual embeddings as an input.
Please help if you have any views on this.

2

Besides just embedding the entire thing (Ada-002 has a massive 8k context), you could just average the embedding vectors, and renormalize by dividing out by the length.

The lossless version is create a matrix of all the embeddings for the conversation.

The fun animated version is have the embeddings along with their time stamps, and resample to a constant smaller unit time interval using Slerp, and project this path to 2D (using PCA?) for human visualization. Maybe one dot per second in the animation.

3

I think that would look super cool, especially with the right dataset.

1 Like