quantizeEmbedding

quantizeEmbedding(v: number[] | Float32Array<ArrayBufferLike>): QuantizedEmbedding

Defined in: src/lib/memoryEngine/quantization.ts:53

Quantize a Float32 embedding (or number[]) into an Int8 vector + scale.

The scale is the maximum absolute value across the input; all other values are mapped linearly into [-127, 127] and rounded. A zero vector yields a zero Int8Array and a scale of 0.

Parameters

Parameter	Type	Description
`v`	`number`[] \| `Float32Array`<`ArrayBufferLike`>	The embedding to quantize. Either a Float32Array (typical for on-device caches) or a plain number[] (typical for values fresh out of `JSON.parse`). Plain numbers are read directly without copying into a Float32Array first.

Returns

QuantizedEmbedding

The quantized data + scale. The returned data.length === v.length.