Some questions about hnsw+rabitq

[suppersam1]

Hi, I’ve integrated HNSW + RaBitQ and SPANN + RaBitQ.
In SPANN, RaBitQ works very well — it can filter out ~90% of candidates, leading to significant speedups.

However, in HNSW, it seems RaBitQ only filters out ~30% of candidates, and the performance improvement is not significant. Moreover, at the same compression rate, the recall with RaBitQ is worse than with SQ.

Has the author tested HNSW + RaBitQ before?
Are there any official benchmarking results or reference data showing how RaBitQ performs when combined with HNSW?

Thanks!

[suppersam1]

This is just my conjecture. The possible reason is that in HNSW, the true top-k results may only be obtained at a relatively late stage of the query, with the early steps only gradually approaching the top-k candidates.In contrast, IVF and SPANN, as clustering-based indexes, start searching from the closest clusters. As a result, the real top-k data can enter the result set much earlier, leading to high filtering efficiency.

[gouyt13]

Thank you for the reported results. We haven't spend much effort on optimizing the HNSW+RaBitQ implementation before. We shall look into this reported issue and improve it soon later.

[suppersam1]

Thank you for the reported results. We haven't spend much effort on optimizing the HNSW+RaBitQ implementation before. We shall look into this reported issue and improve it soon later.

Thanks for your reply. I’m looking forward to your future work and papers.