Best way to change embedding?

[hiepxanh]

there is two case I'm working with Qdrant:

1. I'm changing from embedding model A to model B which have bigger dimension, reason is it's cheaper, or better quality.
2. I'm testing hybrid search using sparse vectors

So the plan is: scan all points, then batch by batch get point value, then push it back:

• step 0: but since collection API is fixed, so I need to create new collection with new vector support. https://qdrant.tech/documentation/concepts/collections/#update-collection-parameters
• step 1: recreate new collection with 2 vectors: https://qdrant.tech/documentation/concepts/collections/#collection-with-multiple-vectors
• step 2: prepare embedding API for sprare vector or new dense vector to embedding
• step 3: batch data to server by: add new point or update vector?

is that a good strategy? do I have any better solution to generate new vector? I would like to have a new article from qdrant team to guide us to generate spare vector or changing new embedding vector from current data

[agourlay]

@hiepxanh Hi 👋

I think your strategy is the right way.

Doing incremental scrolling with transfer into a new collection with batch upload points.

You can play with the scroll/batch size for better performance.

[hiepxanh]

thank you, so much <3