datastax/cassandra used to use the GraphIndexBuilder::rescore method, but I discovered that its design produces suboptimal graphs since it expects a caller to (a) build a graph with one quantization code book then (b) create a new quantization code book and rescore the vectors. The resulting graph produces poor recall with the sift small dataset. Realistically, users should accumulate enough vectors to build a PQ and then insert those vectors using that PQ. I propose that we remove this method entirely to prevent accidental usage of this experimental method. I am removing my usage of the method via datastax/cassandra#2241.
datastax/cassandra used to use the
GraphIndexBuilder::rescoremethod, but I discovered that its design produces suboptimal graphs since it expects a caller to (a) build a graph with one quantization code book then (b) create a new quantization code book and rescore the vectors. The resulting graph produces poor recall with the sift small dataset. Realistically, users should accumulate enough vectors to build a PQ and then insert those vectors using that PQ. I propose that we remove this method entirely to prevent accidental usage of this experimental method. I am removing my usage of the method via datastax/cassandra#2241.