Struct QuantizedVector 

pub struct QuantizedVector {
    pub data: Vec<i8>,
    pub params: QuantizationParams,
    pub norm: f32,
}

Unstable: INT8 quantized vector; struct layout and invariants may change.

Quantized int8 vector with its parameters.

Fields

data: Vec<i8>

Unstable: raw quantized data; invariant ([-127, 127]) enforced by constructor.

Invariant

All values must be in the range [-127, 127]. The value -128 causes incorrect results in AVX-512 VNNI and AVX2 SIMD paths due to vpabsb saturation behavior. The from_f32 constructor enforces this via clamping.

params: QuantizationParams

Unstable: quantization parameters; may be separated from the vector.

norm: f32

Unstable: L2 norm; may be removed or moved.

Implementations

impl QuantizedVector

pub fn from_f32(vector: &[f32]) -> Self

Unstable: quantization constructor; clamping behavior may change.

pub fn to_f32(&self) -> Vec<f32>

Unstable: dequantization; output precision may change with scheme update.

Precision

INT8 symmetric quantization maps [-max_abs, max_abs] to [-127, 127], so the quantization step size is max_abs / 127. The maximum per-element round-trip error is bounded by half a quantization step: max_abs / 254.

For a 384-dim unit-norm embedding (max_abs ≈ 1.0), expect element-wise absolute error ≤ 0.004 and cosine-similarity error ≤ 0.5%.

pub fn dot_product(&self, other: &QuantizedVector) -> f32

Unstable: delegates to dot_product_i8; SIMD dispatch may change.

pub fn cosine_similarity(&self, other: &QuantizedVector) -> f32

Unstable: delegates to cosine_similarity_i8; SIMD dispatch may change.

Trait Implementations

impl Clone for QuantizedVector

fn clone(&self) -> QuantizedVector

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for QuantizedVector

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.