Experiencing High Latency Search Using IVF and HNSW on CPU

[iamsabhoho]

Hello!

I have a dataset of 15000 vectors, with embedding size of 512 float32.

• What are the best parameters for IVF and HNSW index for lowest latency?
• What are the best parameters for IVF and HNSW index for highest accuracy?

Note: I'm already running FLAT index and that looks good!

I wanted to compare the index performances. Thank you!

[junjieqi]

Hi @iamsabhoho, there is no specific number for this question. I think you could benefit from tuning the index with different parameter setup. For example, IVF, you could try to use different number of coarse quantizer, and also you need to tune the nprobe in the search time as well. Similar for HNSW, there is e-efConstruction and efSearch parameters. Here is some example you could take a look and you could do your own parameter tuning based on it https://gist.github.com/mdouze/bf31af9b8d1fa5e0009146f00f21874a?fbclid=IwZXh0bgNhZW0CMTEAAR0751rGGP_b9zEFB5bCvQwfJIcgw1IwrndPCrnMz0O4AbJ-6y8pjpWukso_aem_c_3JCTic1jnkXRY_qGeXdg