Our current assumption is that models often produce a 1v1 mapping to embeddings but that is not the case .... for several models e.g insightface, a single input can produce N embeddings (depending on how many faces are in the image)
Short of duplicating the metadata (which seems like a very bad idea as that would be O(N) for space usage within the DB), we have no apparent way at the moment to make this play nicely with the DB.
Our current assumption is that models often produce a 1v1 mapping to embeddings but that is not the case .... for several models e.g insightface, a single input can produce N embeddings (depending on how many faces are in the image)
Short of duplicating the metadata (which seems like a very bad idea as that would be O(N) for space usage within the DB), we have no apparent way at the moment to make this play nicely with the DB.