Slow and variable vector search performance

Wojciech_Pachowiak · 2024-02-23T15:06:42.628Z

Hi!
I have 2 collections, both with Search Indexes. Documents in both collections contain an OpenAI embedding with 1536 values. The first collection has 10k documents, the second one has 100k documents.

Cluster specs:

Frankfurt
M10 (General)
3 replica nodes
no search nodes

I have 2 issues:

queries have a very variable response times - sometimes it’s as little as 0.5 sec, sometimes it can even reach 4 seconds (especially in the case of the 100k docs collection)
it seems like the response times expierience “cold start” - first queries take the longest, the subsequent ones are pretty ok

Can I do something about it? Can it be sped up? I really would like to have a consistent 0.5-1.0sec repsonse times.

Camillo_N_A · 2024-03-25T22:10:39.258Z

I second this, vector search with 100k records in an M10, it takes 3 - 6 sec to return the result using the vector index

how can I speed it up?

Munna_Ram · 2024-08-08T06:56:06.813Z

I am facing the same issue.

[MOTOR-1350] Variable and slow performance on vector search. - MongoDB Jira

Vivekanandan_Sakthivelu · 2024-08-21T01:26:43.198Z

@Wojciech_Pachowiak

Are you running a dedicated search node ?

MongoDB

Search Nodes Now in General Availability: Performance at Scale with Dedicated...

Search Nodes, now in Public Preview, provide higher availability and better performance at scale with dedicated infrastructure.

If not , your search is within the atlas node . Spinning up a dedicated node can help with performance requirements as search request are routed to this new nodes .