dynvec

Vector storage + HNSW ANN index engine for the Dynomite Rust port.

Node-local persistence via Noxu DB (or an in-memory backend for tests / embedded use).
Two compression codecs: per-vector int8 quantisation (default, 4x compression, ~1% reconstruction error) and IEEE 754 half-precision floats (2x compression, ~0.05% error).
Three distance metrics: Euclidean, cosine, and (negated) dot product.
HNSW approximate-nearest-neighbour index, hand-rolled in this crate -- see docs/dynvecdb/architecture.md for the rationale.
Per-table Engine handle for embedders.

This crate is the engine layer. The primary protocol surface lives in dynomite::vector, which exposes the engine through the existing Redis listener using Redis-Stack-style RediSearch FT.* commands. See docs/dynvec/fold-into-redis-path.md for the architecture.

The standalone HTTP API is retained as a debug / inspection facility and is gated behind the http feature; it is not the primary protocol.

Example

cargo run -p dynvec --example quickstart --features http

In another terminal:

curl -s -X POST localhost:21900/tables \
  -H 'content-type: application/json' \
  -d '{"name":"demo","dim":3,"codec":"int8_quantized","distance":"cosine"}'

curl -s -X POST localhost:21900/tables/demo/vectors \
  -H 'content-type: application/json' \
  -d '{"key":"hello","vector":[1.0,0.0,0.0],"metadata":{"label":"x"}}'

curl -s -X POST localhost:21900/tables/demo/search \
  -H 'content-type: application/json' \
  -d '{"vector":[0.95,0.05,0.0],"k":3}'

dynomite-vec 0.0.1

dynvec

Example