Hi Sergei,

I first tried M=15 and EF=50 then M=30 and EF=100 and the latter
appears to work better search results wise - I haven't
measured/compared perf against itself or Chroma yet.

Thanks again!

On Thu, 30 Jan 2025 at 17:20, Sergei Golubchik <[email protected]> wrote:
>
> Hi, Vishal,
>
> Better to observe a bit. I'd expect different HNSW implementations to
> behave somewhat similarly, but not identically. That is, your values
> from Chroma could be a good starting point, but might need some
> tweaking.
>
> MariaDB uses a modified version of HNSW that generally produces better
> recall than the original algorithm. This difference becomes larger as M
> inceases.
>
> What M values do you use?
>
> On Jan 30, Vishal Rao via discuss wrote:
> > Hello,
> >
> > I'm replacing the use of Chroma with MariaDB as the vector store and
> > the Chroma code sets the ef_search parameter to a value of 50.
> >
> > My question is, to achieve "similar results" with the MariaDB vector
> > store will setting ef_search to 50 here as well have the intended
> > effect?
> >
> > Or is there no relation between the actual values across
> > implementations and I need to experiment/test to observe how MariaDB
> > performs to arrive at an appropriate value for it?
> >
> > Thank you.
> >
> Regards,
> Sergei
> Chief Architect, MariaDB Server
> and [email protected]



-- 
"The World is a book, and those who do not travel read only a page." -
St. Augustine.
_______________________________________________
discuss mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to