Hacker News Reader

Comment by rcarmo | Hacker News Reader

Well, I see this as great news. I've been tinkering with vector search myself and really don't like the APIs, query structures, even the embedding steps required. Anything that makes it simpler (and faster) is most welcome.

I do hope Salvatore keeps doing emulators for MCUs, though :)

View on HN

Thank you Rui, my target is: API that can be learned and used under 120 seconds. I don't think more of this is needed, for the basic usage. Then there could be other stuff, but it must be incremental. Easy things must be easy.

Well, the VSIM example I got inside of 10 seconds, so at least for now you're on track :)

TLDR API:

VADD key [VALUES <float float ...> | FP32 <binary blob>] element

VSIM key [VALUES ... | FP32 ... | ELE ...] [COUNT <count>] [EPSION <max distance>]

VCARD key (number of elements)

VDIM key (dimension of the vectors)

VDEL key <ele>

And so forth. However there are nice things you can enable, if you want. For instance VADD has the [REDUCE] option that implements random projection in order to reduce the dimensionality of your embeddings. Less memory, and it is faster. And so forth.

Will work on it during xmas, it's quite fun.

It implements on-insertion normalization and int8 quantization. The algorithm to insert nodes inside the HNSW attempts to never leave isolated nodes and to get good quality links and so forth. More hacking needed.

Huh Charles actually can code? We’ll then.

yes loving this! please PLEASE be the hero we need in unfucking the unholy mess of misdirection and useless abstraction that is storing some vectors somewhere for similarity search right now

U mean there are too much leaky abstractions in the field now ?

chroma's api i think takes something like 20 seconds

Reply on news.ycombinator.com