• added db locked multi-instance support with dynamic paths. multiple instances with different databases will work, multiple instances with same db name will fail
• gpu and cuda arch check on start to prevent crashing.
• sdk updated

New features:

• Vector hashing to prevent duplicate records
• "Exists" endpoint to check is given vector exists on db

Due to performance concerns all text mode communication is removed. use sdk or take a look to datasheet.md for custom implementations.

• full protocol overhaul
• clustering with single command
• multi-db deployment with "vec deploy"
• extra payload support for small blobs under 100kb.

binary protocol support for router mode

fixed router issues and added avx instructions to take advantage of modern cpu's in vec-cpu version

created cpu version and router mode. single port, 5 instances. commands works with an extra database prefix

 # Start multiple DBs without TCP
  vec --notcp tools 1024
  vec --notcp conversations 768
  vec --notcp embeddings 512

  # One router exposes them all on port 1920
  vec --route 1920

  Then from a client, append db names next to commands:

  pull tools 0.1,0.2,...
  cpull conversations 0.3,0.4,...

  The router parses the first word as the instance name and forwards the rest to the matching pipe/socket.

What's New

box utility is now an internal part of vec. an attached label alongside
vectors returns back on queries. it can be a single word or a uri scheme.

push path/to/file/test.pdf?page=3 0.23231,0.424445 -> 42
push file.jpg?person=1&w=1024&h=1024 0.23231,0.424445 -> 43

pull 0.23231,0.424445,...
-> path/to/file/test.pdf?page=3:0.001234,file.jpg?person=1&w=1024&h=1024:0.045100

new commands: bpull, bcpull (binary queries), label (set/override)
box project removed. single exe does everything.

• added cosine angle query function (cpull command)
• fp32 is default now, fp16 offered optionally
• created companion app "BOX" a string metadata store in same logic of "VEC"

added ada and turing architecture targets