Helpers¶

Lower-level building blocks exposed for advanced use. Most users should reach for one of the index classes.

`rht`¶

rht(x: NDArray[float32], seed: int) -> NDArray[np.float32]

Randomized Hadamard Transform: D·H·x / sqrt(d).

Parameters:

Name	Type	Description	Default
`x`	`(NDArray[float32], shape(..., d))`	Input vector(s). `d` must already be a power of 2.	required
`seed`	`int`	Rotation seed. Use the same seed consistently for an index.	required

Returns:

Type	Description
`NDArray[np.float32], same shape as x.`

padded_dim(dim: int) -> int

Return the WHT-padded dimension for a given embedding dimension.

get_codebook(bits: int) -> tuple[NDArray[np.float32], NDArray[np.float32]]

Return (centroids, boundaries) for the given bit-width.

Parameters:

Name	Type	Description	Default
`bits`	`int`	Quantization bits. Must be 2, 3, or 4.	required

Returns:

Name	Type	Description
`centroids`	`(NDArray[float32], shape(2 ** bits))`	Reconstruction values.
`boundaries`	`(NDArray[float32], shape(2 ** bits - 1))`	Decision thresholds for np.searchsorted.