In order to apply a model beyond the initial data used for exploration, the anchors need some way to tell if the data they are featurizing on is new or not. This primarily applies to similarity anchors, if they've cached any embedding vectors or similar on the anchorlist, they won't know that with new data they need to recompute them.
Perhaps keep a hash of the data that gets sent through featurize that the anchors could check for.
In order to apply a model beyond the initial data used for exploration, the anchors need some way to tell if the data they are featurizing on is new or not. This primarily applies to similarity anchors, if they've cached any embedding vectors or similar on the anchorlist, they won't know that with new data they need to recompute them.
Perhaps keep a hash of the data that gets sent through featurize that the anchors could check for.