Hi!

The implementation of the HNSW algorithm in MariaDB is not identical to the paper (hence the m from modified or MariaDB in mHNSW).

You should not try to compare against equal M / ef_search values of other implementations, but rather compare QPS for a particular Recall target (say 95%, 98%, 99%)

That will give you an apples-to-apples comparison.

Vicențiu

On Wed, 19 Feb 2025 at 14:46, shiyer22--- via discuss <discuss@lists.mariadb.org> wrote:
Thanks Sergei! Is it possible to let me know the internally decided ef_construction value? That will help me in putting recall results in context with other benchmarks. Is the ef_construction computed based on M / dimension  etc?
_______________________________________________
discuss mailing list -- discuss@lists.mariadb.org
To unsubscribe send an email to discuss-leave@lists.mariadb.org