Convert to CSR format for efficient row access. The graph is already CSR, so this is a copy into the standalone [`CsrMatrix`] type.

Extract edges as (head, tail, weight) arrays for the optimizer, in row-major / column-sorted order.

Whether to use RP tree initialization. Default: true.

Uniform-buffer bind group layout entry.

Perform a vector search with the given query

Enable verbose output. Default: false.

Enable verbose output. Default: false.

Create an Embedding Atlas widget. Args: data_frame: A DataFrame/Arrow object to "register" with DuckDB.

Return an async embedder backed by LiteLLM.

Return an async embedder backed by SentenceTransformers (text only).

Return an async embedder backed by CLAP for audio data.

Return an async embedder backed by HuggingFace transformers pipelines.

Return an async embedder backed by a HuggingFace image-feature-extraction pipeline.

Return an async embedder backed by a HuggingFace feature-extraction pipeline.

Return deterministic random vectors seeded by batch length.

Bit-packed versions that work with &[u8] directly

Create a builder for UMAP dimensionality reduction. # Arguments `data` - Array of shape (n_samples, n_features)

Create a builder for a new NNDescent index. # Arguments `data` - Array of shape (n_samples, n_features) `metric` - Distance metric name (e.g., "eucli

Create a builder from a precomputed kNN graph, skipping NNDescent. No high-dimensional data is needed — once the graph exists, UMAP uses only it. `kn

Query neighbors for all indexed points in one call, avoiding per-point worker round-trips.

Provide a fresh temporary cache directory and clear the constants LRU cache.

Compute dot-product projections of points onto a hyperplane. Returns `margins[i] = dot(hyperplane, data[indices[i]]) + offset` for each index. Used b