Integrate BBC with graph-quantization ANN methods

Develop an integration of the bucket-based result collector (BBC) with graph–quantization approximate nearest neighbor methods that combine quantization and graphs, such as SymphonyQG and NGT-QG, to extend BBC beyond pure quantization-based IVF indexes.

Background

The paper introduces BBC, a bucket-based result collector designed to accelerate quantization-based approximate nearest neighbor (ANN) methods for large-k queries. While BBC is integrated and evaluated with IVF+PQ and IVF+RaBitQ, the authors note that there also exist ANN approaches that integrate quantization with graph structures.

However, the authors report practical limitations in existing graph–quantization systems: SymphonyQG and NGT-QG exhibit out-of-memory issues during indexing on their setup due to high memory consumption, and LVQ is closed-source. Given these constraints, the authors explicitly leave integrating BBC with such graph–quantization methods as future work.