Store Only PQ Vectors Without Original Embeddings & Update Non-Vector Fields in OpenSearch KNN

OpenSearch
1

Hi all,

I’m trying to configure a large-scale OpenSearch KNN index with the following requirements:

  1. No original embeddings on disk – only PQ-compressed vectors for ANN search.

  2. Update non-vector fields without re-uploading or re-calculating vectors.

  3. No need to return original embeddings in search results.

I’d like advice on the best way to set _source, derived_source, or knn_vector settings in OpenSearch 3.x to fully support this workflow.

Any examples or guidance would be greatly appreciated!

Thanks!