High-performance FISHDBC clustering in Rust.

Chutoro implements the FISHDBC algorithm — a scalable, density-based clustering method that discovers clusters of arbitrary shape without requiring the user to specify the number of clusters or a distance scale. It builds on the lineage of DBSCAN and HDBSCAN* while replacing the expensive all-pairs distance matrix with an approximate HNSW graph, making it practical for large and high-dimensional datasets.

• No parameter guessing: FISHDBC extracts an optimal flat clustering from a density hierarchy — no need to choose eps or k up front.
• Bring your own metric: implement the DataSource trait with any distance function, and chutoro handles the rest.
• Rust-safe parallelism: the CPU backend uses Rayon for parallel HNSW insertion and MST construction, with zero unsafe in user-facing code.
• Extensible by design: feature-gated backends (CPU today, GPU planned) and a builder API that stays stable as the engine evolves.

[dependencies]
chutoro-core = "0.1.0"

use chutoro_core::{
    ChutoroBuilder, DataSource, DataSourceError, ExecutionStrategy,
};

/// Each row is one point in 2-D space.
struct Points(Vec<[f32; 2]>);

impl DataSource for Points {
    fn len(&self) -> usize { self.0.len() }
    fn name(&self) -> &str { "points-2d" }
    fn distance(&self, i: usize, j: usize)
        -> Result<f32, DataSourceError>
    {
        let a = self.0.get(i)
            .ok_or(DataSourceError::OutOfBounds { index: i })?;
        let b = self.0.get(j)
            .ok_or(DataSourceError::OutOfBounds { index: j })?;
        Ok(a.iter()
            .zip(b)
            .map(|(x, y)| (x - y).powi(2))
            .sum::<f32>()
            .sqrt())
    }
}

let chutoro = ChutoroBuilder::new()
    .with_min_cluster_size(3)
    .with_execution_strategy(ExecutionStrategy::CpuOnly)
    .build()?;

let result = chutoro.run(&Points(vec![
    [0.0, 0.0], [0.1, 0.1], [0.2, 0.0],   // cluster A
    [5.0, 5.0], [5.1, 4.9], [5.0, 5.2],   // cluster B
]))?;

println!("Found {} clusters", result.cluster_count());
# Ok::<(), chutoro_core::ChutoroError>(())

• Four-stage CPU pipeline: HNSW construction, candidate edge harvest, Kruskal MST, and stability-based hierarchy extraction (design document).
• Built-in Euclidean and cosine distance helpers with input validation and pre-computed norm caching (users' guide § distance helpers).
• Property-tested (proptest suites) and formally verified — Kani harnesses for HNSW graph invariants and distance-metric properties; Verus proofs for edge harvest ordering and extraction (developers' guide § Verus proofs).
• Criterion benchmarks with synthetic data generators (Gaussian blobs, ring/Swiss-roll manifolds, text mutation, MNIST download) (developers' guide § benchmarks).
• Optional metrics crate integration for distance-cache telemetry (users' guide § feature flags).
• CLI tool (chutoro-cli) and bundled data-source providers: dense vectors via Parquet/Arrow (chutoro-providers-dense) and text via Levenshtein distance (chutoro-providers-text).

• Users' guide — full API documentation and usage patterns
• Developers' guide — contributing, benchmarks, and formal verification
• Design document — architecture and literature survey
• Roadmap — planned features and progress

ISC — see LICENSE for details.

Contributions welcome! Please see AGENTS.md for guidelines.