Gets an object from the cache with the given key. Returns `null` if the entry is not found.

Sets an object to the cache with the given key

Compute embeddings and generate 2D projections for a DataFrame column. This is a unified entry point that auto-detects the modality of the i

Compute PageRank scores from a list of edges of a graph using PyTorch sparse matrix power iteration. The graph can be either unweighted (each

Add data to the summarizer

Creates a derived value.

Create a narwhals Series from a list of values via pandas.

Random seed for reproducibility.

Run UMAP and return the result (embedding + KNN graph).

Build the NNDescent index. This is a synchronous wrapper around [`build_async`]. On native builds (including native GPU via `pollster::block_on`), us

Random seed for reproducibility.

Detect whether binary data is an image or audio based on magic bytes.

Number of training epochs. Default: auto.

Creates a stateful derived value.

A fast (pseudo)-random number generator. Must exactly match the Python bitmask arithmetic. The Python code uses i64 state but performs u32-level bit

Map the inputs by an async function, return a future that resolves to the mapped array (in correct order). Args: inputs: List of ite

Initialization method. Default: Spectral.

Number of nearest neighbors. Default: 15.

Store a value in the encrypted file cache. This function serializes, encrypts, and stores a value in the file cache using the provided k

Get the neighbor graph (indices, distances). Applies distance correction if using fast alternative distances.

Infer the modality by inspecting the first non-null value in the series.

Generate test data matching conftest.py nn_data fixture. Shape: (1002, 5) - 1000 random rows + 2 zero rows. Uses our TauRng for deterministic generati

Helper: assert two f32 values are approximately equal.

Print a message to stderr if verbose is enabled.

Destroy the instance and release resources

Retrieve a cached value from the encrypted file cache. This function attempts to load and decrypt a previously cached value using the pr

A fast (pseudo)-random number generator for floats in [0, 1].

Fill the given rect as occupied

* Anneal the layout to completion: linear learning-rate decay from the current * `learningRate` down to 0 over the default horizon, continuing from

* Advance the layout by `nEpochs` at the current `learningRate`, held flat (no * decay). Intended for interactive/real-time stepping. Default: 1 ep

Compute recall: fraction of true neighbors found.

Build the graph + initial embedding and return a resumable [`UmapSetup`], without** running the layout optimization. Use the returned [`UmapOptimizer`

Convert series to canonical image format: list[{"bytes": bytes}].

Push a value with a flag into a max-heap, checking for duplicates. Returns 1 if inserted, 0 if rejected or duplicate.

Enable GPU acceleration for nearest neighbor computation. Default: false. Requires the `gpu` crate feature. Falls back to CPU if no suitable GPU is a

Enable GPU acceleration for distance computation. Default: false. Requires the `gpu` crate feature. When enabled, distance computation is offloaded t

Convert raw UMAP k-nearest-neighbor (KNN) arrays into a weighted edge list, which can then be passed into pagerank(). Raw KNN distances

Query the index for k nearest neighbors. @param query_data - Flat Float32Array (n_queries * n_features), row-major. @param n_queries - Number of quer

Query the index for k nearest neighbors of query points. # Arguments `query_data` - Array of shape (n_queries, n_features) `k` - Number of neighbors

Async version of ``file_cache_value``. Identical behaviour but *value_func* is awaited on a cache miss.

The current embedding (CPU mirror; valid after `step`/`run_to`).

Retrieve a cached value or compute and cache it if not present. This is a read-through cache helper: if a cached value exists for the given

Find a start location (x, y) with a rect of size (width, height) can be placed

Generate deterministic spatial test data matching conftest.py. Shape: (12, 20) - 10 random rows (seeded) + 2 zero rows. We use our own TauRng for repr

Canonicalize a precomputed kNN graph into the form [`fuzzy_simplicial_set`] expects: column 0 is the point itself (distance 0) and the remaining colum

Pop (exit) the current stage. Reports progress at stage end.

Push (enter) a stage with a custom verbose message. Reports progress at stage start and logs the message.

* Invoke the method on target if defined, and if all args are defined.

Convert series to canonical text format: list[str] with nulls as 'null'.

Convert series to canonical vector format: list[ndarray[float32]].

Brute-force k-nearest neighbor for query points against a data set.