What was the selling point to use arroy

[irevoire]

Hello team, co-author of arroy here!

I just discovered your project through this blog post, and it’s really awesome to see this kind of project using our vector store 🔥

I had a few questions about it:

• What were you using before if there was something?
• What was the main argument to switch to arroy over something else?
• Have you seen any perf improvement upon making the switch? 👀

[ealmloff]

Hello team, co-author of arroy here!

I just discovered your project through this blog post, and it’s really awesome to see this kind of project using our vector store 🔥

Hello, thanks for your work on arroy, it has been great to work with so far!

I had a few questions about it:

• What were you using before if there was something?

I was using instant-distance before switching to arroy.

• What was the main argument to switch to arroy over something else?
• Have you seen any perf improvement upon making the switch? 👀

The reason I switched to arroy was to make incremental indexing possible. I haven't done any formal benchmarks before and after the switch, but I am hopeful that incremental indexing will speed up workloads with a constantly updating context significantly. For example, this code listens to your environment, transcribes and embeds text and uses arroy to find context that is similar to questions. As the context grows, performing a full rebuild of the index would not be feasible

[irevoire]

Niiiice, on my side, I tried to measure the worst-case scenario for the incremental indexing and got a 10x speedup when arroy needed to create a new tree and something like a 30-40x speedup when we didn’t need to create a new tree.
You’ll probably get something even better with a normal use-case instead of the worst case scenario