Crate precinct 

Approximate nearest-neighbor search over region embeddings (boxes, balls).

Point-based ANN indices (HNSW, IVF, Vamana) assume queries and database entries are single vectors. Region embeddings – axis-aligned boxes, balls, cones – represent concepts as volumes in embedding space. precinct bridges this gap.

Core abstractions

• Region – trait for geometric regions with center, point-to-region distance, and containment.
• AxisBox, Ball – concrete region types.
• RegionIndex – ANN index over regions. Builds a point index over region centers, retrieves candidates, reranks with true region distance.

Usage

use precinct::{AxisBox, Region};

let b = AxisBox::new(vec![0.0, 0.0], vec![1.0, 1.0]);
assert!(b.contains(&[0.5, 0.5]));
assert!(!b.contains(&[1.5, 0.5]));
assert_eq!(b.distance_to_point(&[0.5, 0.5]), 0.0);

Re-exports

pub use index::Error;

pub use index::IndexParams;

pub use index::RegionIndex;

pub use index::SearchParams;

Modules

index

io

Simple binary I/O for region embeddings.

Structs

AxisBox

Axis-aligned hyperrectangle defined by min/max corners.

Ball

Hypersphere defined by center and radius.

Traits

Region

A geometric region in d-dimensional space.

Functions

ball_to_point_l2

L2 distance from a point to a ball (hypersphere) surface.

box_to_point_l2

L2 distance from a point to an axis-aligned box.