Different results when doing a query with or without and HNSW index about pgvector HOT 5 CLOSED

Hi @jpbalarini, you'll see different results for queries that use an approximate index (docs). You can try increasing ef_construction (and then m if that doesn't work) to see if it improves recall.

Do you have a lot of duplicate embeddings? It's possible that could be contributing to poor recall.

from pgvector.

Just noticed you're using m = 4 and ef_construction = 10. This is likely why recall is poor. Try using the defaults.

from pgvector.

Thanks @ankane . I found where I took those parameters from:
https://www.crunchydata.com/blog/hnsw-indexes-with-postgres-and-pgvector#tldr
No idea why they use those values.

Leaving the default values (partially) did the trick. I now get something that makes some sense, but I'm not getting the values that I need that are closer (0.84 and 0.82).

Without using the index (seq scan):

Using the index:

I tried increasing the ef_construction from 64 to 96, but the results were the same. I'm now trying with 128 (still building the index). Do you have something to recommend? Maybe trying to increase the m size too?
Any recommendation is useful; thanks!

from pgvector.

I'd recommend calculating recall programmatically over many queries to quantify the difference between sets of parameters. The best parameters will be different for each dataset. I'd try ef_construction = 64, 128, 256, and 512 (with m = 16) and then m = 32 and 64 (with ef_construction being at least 4 * m) until you reach your target recall.

from pgvector.

Thanks Andrew for your insights! I will try that 💪🏻

from pgvector.