Recursively build an angular RP tree (post-order node storage).

Generate deterministic binary test data. Shape: (12, 20) - 10 random rows + 2 zero rows.

Compute epochs_per_sample for each edge based on weight. Edges with higher weight are sampled more frequently.

Recursively build a euclidean RP tree.

Recursively build a hub-based angular tree.

Recursively build a hub-based euclidean tree.

Creates a server for hosting Embedding Atlas

Async helper: map a buffer for reading and wait for it to be available.

Largest edge weight (NEG_INFINITY if empty).

Minimum distance between points in embedding. Default: 0.1.

Multiply by a small dense matrix: result = self * coeffs. coeffs is (self.ncols() × k_out), row-major.

Multiply this matrix by a dense f64 vector, accumulating in f64. The matrix data (f32) is promoted to f64 for the computation.

Number of random projection trees. Default: auto.

Get the neighbor indices for a vertex.

Build new and old candidate neighbor lists from the current graph. Returns (new_candidate_indices, old_candidate_indices), each shape (n_vertices, max

The internal NN-descent iteration loop.

Compute the L2 norm of a vector

Orthonormalize self against a basis block, then self-orthonormalize. self -= basis * (basis^T * self), then MGS-orthonormalize self.

Prepare the search index for querying.

Prepare the search graph for querying. This must be called before `query()`, or `query()` will call it automatically. # Panics Panics if the neighbo

Print a formatted summary table.

Print a formatted summary table.

Process candidates: fused distance computation and heap update. Phase 1: Each thread processes its own block of candidate rows, pushes into its own he

GPU-accelerated candidate processing. Snapshots heap thresholds, collects candidate pairs, dispatches distance computation with threshold filtering t

Set a progress callback. Default: none.

Set an optional progress callback. Default: none.

Remove entries below threshold.