Holy i didn't realize you guys were working on allowing for vector storage as well. This is hype team. Can you give any guidance on what to expect with the various typesense cluster sizes & how much memory the vectors take up in terms of space?

It would depend on the number of dimensions in your embeddings. These are essentially floating point arrays, so if you have X dimensions per embedding, Y number of documents, you’d need at least X * Y * 4 * 2 bytes of RAM to index the embeddings

filter_by

still works tho?

Yup!

Damn this is wild. We're gonna have to move our vectors off pinecone and onto typesense now too.

Out of curiosity, does PineCone support text keyword queries with vector search?

Nope. Just vector search. We filter by the metadata after matching K closest vectors.

I see

Hence why keyword + vector search would be a game changer.

Btw, does pinecone support filter_by with vector search?

Yeah it does Jason. You can filter by "metadata" which is indexed. It's roughly 500k 1536 dim floating point vectors I can store (with only 1kb of metadata associated w/ the vector) per "pod" as they call it. Which is ~ $0.10/hr on pinecone. So the comparable typesense cluster at 8gb at $0.17/hr can store roughly the same amount of vectors Based on your math, I can store 500k vectors * 1536 * 8 = 6.14gb with another 2gb to spare for metadata (roughly 4kb per vector)

Thanks Robert! That’s really good context to have.