How to use customized embeddings from OpenAI Cookbook?

Hello Everyone,

Context :
I am following this cookbook from OpenAI for customizing embeddings :

Here is my usecase :
We have huge number of private documents (~400000) which is filled with organization specific words and knowledge.
We have already divided documents into chunks and converted each chunk to embedding and stored it in a VectorStore (OpenSearch).
But while retrieving a lot of irrelevant documents are found similar to query.
In order to circumvent that, we are planning to create bias matrix as described in OpenAI Cookbook for customiziing embeddings.

Question :
The bias matrix (it is mentioned as best matrix in code) generated as part of the cookbook is of shape : (1536,2048)
The embedding generated using ADA model has shape : (1,1536)
What should be my precise next steps ?

Possible Next Steps :
Step 1 : Multiple emebedding of every chunk (which is shape of (1,1536) with bias matrix (which is shape of (1536,2048)) that results into matrix of shape (1,2048).

Step 2 : Insert result of multiplication (i.e. (1,2048) matrix) to vector store embedding against chunk.

Step 3 : While querying vector store, follow the Step 1 and Step 2 again for embedding generated for query

Hi! Welcome to the forum!

Yep, you’ll have to run the transformation every time you get your embeddings from the API.

Is that your question?