Scores rows of a packed store against a query and ranks them.
With a numeric TopK this keeps a bounded, stable insertion buffer, so the
cost is O(rows × dims) with no full sort. With TopK: null it returns every
row that passes the threshold, fully sorted. NaN scores (only possible from
NaN or infinite vector values) are never returned.
Scores rows of a packed store against a query and ranks them.
With a numeric
TopKthis keeps a bounded, stable insertion buffer, so the cost is O(rows × dims) with no full sort. WithTopK: nullit returns every row that passes the threshold, fully sorted. NaN scores (only possible from NaN or infinite vector values) are never returned.