The idea here is to extend, for internal search use, the data structure of a SBF to hold for now a few vector sizes
vector_576 (yolov8)
vector_1024(mobileNetv3)
vector_512(Apple Vision ML Image Finger Print, I think this is efficientNet ...)
vector_384(Bert, text embeddings. has a hidden size of 384, Sorry had forgotten this one)
ML_metadata (info about the model that generated the embeddings in JSON format)
Object detection/labels will go into current elements and for Object detection we will (guess what) use miniOCR with values containing both the Label + the confidence... so people can't use this directly for exposing to the world @alliomeria
The limit is 1024 because that is a default performance option in Solr. Could be raised but nobody in our community can afford a super computer to ask for image similarity.
What?
The idea here is to extend, for internal search use, the data structure of a SBF to hold for now a few vector sizes
Object detection/labels will go into current elements and for Object detection we will (guess what) use miniOCR with values containing both the Label + the confidence... so people can't use this directly for exposing to the world @alliomeria
The limit is 1024 because that is a default performance option in Solr. Could be raised but nobody in our community can afford a super computer to ask for image similarity.