Pros and cons of vector search in elastic?

Posted by ernestdong on Wednesday July 19, 2023 @08:34AM
from the self-post dept.

I'm comparing Elastic vs other pure vector databases vs Mongodb/redis offerings. Is anything wrong or supplemental? Thank you!

Pros: - It's an Elastic product, meaning high SLA and needless to buy other products when doing business with Elastic. - Have built-in Embedding models: ELSER. - Hybrid search with text+vector - Security

Cons: - [Doesn't support quantization](https://discuss.elastic.co/t/about-the-dense-vector-compression/319398), which may be too exhaustive when storing vast numbers of vectors. - [Performance seems poor](https://ann-benchmarks.com/) - Does not support CUDA/GPU, versus milvus support Nvidia RAPIDS

( 2 comments | on HN )

Re:Pros and cons of vector search in elastic? by andre-z on Wednesday July 19, 2023 @08:49AM (#36785464)

It depends on your requirements, for a simple hybrid-search solution, elastic, etc. should be enough. With growing data amount and if working not only with text embeddings, you should try out a dedicated solution, like Qdrant. https://github.com/qdrant/qdrant (Disclaimer, I'm from the team)

[ Parent | on HN ]

Re:Pros and cons of vector search in elastic? by philipvollet on Wednesday July 19, 2023 @09:18AM (#36785812)

Highly opinionated as I'm working for Weaviate, so take my comment with a large portion of salt.

As always depends on your use-case.

My highly opinionated view is that for Elastic, they're not really open source and the dependency on Java of the Lucene ecosystem is a big disadvantage, so as you already said, speed, they're getting better at this, but if you need to scale, this problem scales with you.

So if you already have ELK stack and don't need to scale, sure go for it otherwise, Weaviate offers real open source, so use it for free on your own infrastructure https://github.com/weaviate/weaviate

[ Parent | on HN ]