Find answers from the community

Updated 2 years ago

What is the fastest way to build vector

At a glance

What is the fastest way to build vector index from nodes with precomputed embeddings?

Is it just index = VectorStoreIndex(nodes)?

It still seems to take a long time (20mins for 400k nodes), considering that the embeddings for nodes are already computed.

12 comments

LLogan M

So every node has node.embedding already set?

DDove

Yes

LLogan M

hmm ok good. Although I just read your message again and you have 400K nodes lol

LLogan M

at that point, loading that all into memory and moving stuff around is going to be the biggest bottleneck

LLogan M

not much else to do to improve speeds

LLogan M

with that many nodes I highly recommend using a vector DB integration

DDove

Which DB do you recommend?

LLogan M

weaviate or pinecone is nice for hosted stuff

Self-hosted I would use qdrant or chroma

Tbh I think qdrant will be the fastest

DDove

Got it, thanks!

LLogan M

Like, it might still be slow-ish to insert 400K nodes, but that should be a one time cost 🙂

DDove

I tested the :memory: mode for qdrant, and it seemed to be sligtly faster than the vanilla in-memory storage

LLogan M

I'd you can run a hosted version, like with their docker image, it may be faster too (pretty sure the backend is all written in rust then)

Add a reply