Method random_layer

nodedb-vector/src/hnsw/graph/index.rs:348–353

Assign a random layer using the exponential distribution. Capped at `MAX_LAYER_CAP` to prevent pathological RNG draws from promoting the index's `max_layer` to hundreds or thousands, which would make every search's Phase-1 greedy descent O(max_layer).

(&mut self)

    /// promoting the index's `max_layer` to hundreds or thousands, which
    /// would make every search's Phase-1 greedy descent O(max_layer).
    pub(crate) fn random_layer(&mut self) -> usize {
        let ml = 1.0 / (self.params.m as f64).ln();
        let r = self.rng.next_f64().max(f64::MIN_POSITIVE);
        let layer = (-r.ln() * ml).floor() as usize;
        layer.min(MAX_LAYER_CAP)
    }

    /// Compute distance between a pre-encoded query and a stored node.
    ///