Phase II · High-Velocity Concurrent Processing Tier

Billion-scale nearest neighbour over distributions, in Rust.

Concurrent HNSW with a pluggable distance metric — and the first production-ready Sliced-Wasserstein implementation that composes cleanly with HNSW. Apache 2.0.

ef_search=50

Click a node to inspect it.

Metric

SW₁

Candidates eval.

Nearest dist.

0.0866

Recall@5

100%

What you get

→vector-index — generic ConcurrentHnsw<P, M> over a parking_lot::RwLock-backed index.
→sliced-wasserstein — SW₁ over point-cloud fingerprints; implements Metric from vector-index behind a feature flag.
→Stable insert/search under high-throughput concurrent ingest.

What it solves

Cosine and L2 collapse semantically rich fingerprints; SW₁ preserves distributional structure.
HNSW gives you sub-millisecond k-NN. Compose the two for a billion-scale, distributionally-aware index.

Pricing

Apache 2.0 — free

Enterprise use cases

Catalogue dedup / matching

Stock-audio and stock-video libraries at multi-billion-asset scale.

AI-content attribution

Find the closest known fingerprint to a freshly generated piece of media.

Recommendation over empirical distributions

When the 'item' is a behavioural histogram, not a point in space.

Quickstart

use vector_index::{HnswConfig, concurrent::ConcurrentHnsw};
use sliced_wasserstein::{PointCloud, SlicedWasserstein, SwConfig};

let metric = SlicedWasserstein::new(SwConfig { dim: 172, n_projections: 100, seed: 42 });
let index  = ConcurrentHnsw::new(HnswConfig::default(), metric)?;
index.insert(point_id, PointCloud::new(fp, 172)?)?;
let hits = index.search(&query, 10);

↗ View on crates.io Star omnipulse-rs on GitHub ↗Talk to founders →