Creating index is very intensive

I noticed that when creating an index, the passages are encoded twice:

- once, when creating the indexing plan
- one more time when doing the actual indexing

Is it possible to save the already encoded vectors, as encoding depending on hardware can take a very long time, especially if done twice for each passage per index creation/addition operation?
Also is it possible to reuse already created BERT embeddings when adding to index?

Start by tracing the index-creation flow and locating where passages are encoded during planning and actual indexing. Determine whether encoded vectors or existing BERT embeddings can be retained and reused for later additions, then verify that indexing avoids duplicate encoding while producing the expected index.

Written by the indexing model from the issue text.