- 
                Notifications
    You must be signed in to change notification settings 
- Fork 25.6k
Open
Open
Copy link
Labels
:Search Relevance/VectorsVector searchVector search>enhancement>featureTeam:Search RelevanceMeta label for the Search Relevance team in ElasticsearchMeta label for the Search Relevance team in Elasticsearch
Description
We need to provide an index time configuration item for enabling DirectIO for all our quantized vector types. Currently, its possible to only enable direct IO through a java feature flag. This of course is difficult for users to enable and it applies to only BBQ vector types and does so indiscriminately.
We must be able to enable DirectIO for the high fidelity vectors provided when quantized search is enabled and reranking must take place.
Since this should be possible for all quantized values, and we want to be able to quantize from bytes, and have bfloat16 vectors, I am not sure a fully qualified format name will be scalable or usable for this.
I think we will need fully composable, pluggable, and backwards compatible formats: #134394
Metadata
Metadata
Assignees
Labels
:Search Relevance/VectorsVector searchVector search>enhancement>featureTeam:Search RelevanceMeta label for the Search Relevance team in ElasticsearchMeta label for the Search Relevance team in Elasticsearch