Advantages over Elasticsearch kNN

[andrasaponyi]

Hi,

I'm learning about vector similarity search in Elasticsearch and have recently come across this plugin as an alternative to the default kNN implementation. I understand that previously Elasticsearch did not support approximate nearest-neighbor search, but aNN was actually implemented in version 8.0.

So I've been wondering - what are the advantages of elastiknn over the native Elasticsearch implementation at this point?

Here are some advantages I can see from what I've read so far:

• elastiknn supports multiple vectors per document
• some similarity metrics not supported by ES (L1, Jaccard, Hamming)
• elastiknn uses Locality Sensitive Hashing as opposed to HNSW (but is this really an advantage?)

Looking forward to your answers 🙂

[alexklibisz]

Hi, I have only used Elasticsearch kNN for some very quick/simple experiments, several months ago. So take this with that grain of salt.

The first two you pointed out are accurate. Elastiknn also supports sparse vectors; AFAIK, elasticsearch does not.

Whether LSH is an advantage I guess depends on the problem and dataset at hand. I think HNSW is better on most benchmarks.

[andrasaponyi]

Thank you, @alexklibisz, that's very helpful!