Struct QuantizedEmbedding 

pub struct QuantizedEmbedding {
    pub values: Vec<i8>,
    pub params: QuantizationParams,
    pub hash: [u8; 32],
}

Int8 quantized embedding with metadata

Achieves 4x memory reduction compared to f32 embeddings. Following Gholami et al. (2022) survey on quantization methods.

Fields

values: Vec<i8>

Quantized values in [-128, 127]

params: QuantizationParams

Quantization parameters for dequantization

hash: [u8; 32]

BLAKE3 content hash for integrity (Poka-Yoke)

Implementations

impl QuantizedEmbedding

pub fn from_f32( embedding: &[f32], calibration: &CalibrationStats, ) -> Result<Self, QuantizationError>

Quantize f32 embedding to int8 with calibration

Implements symmetric quantization: Q(x) = round(x / scale)

Errors

• EmptyEmbedding: If input is empty
• DimensionMismatch: If dimensions don’t match calibration
• NonFiniteValue: If NaN/Inf detected (Poka-Yoke)

pub fn from_f32_uncalibrated( embedding: &[f32], ) -> Result<Self, QuantizationError>

Quantize f32 embedding using local statistics (no calibration)

Computes absmax from the embedding itself.

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

Dequantize to f32 embedding

Returns approximate original values within error bound.

pub fn verify_integrity(&self) -> bool

Verify integrity hash (Poka-Yoke)

pub fn dims(&self) -> usize

Get embedding dimensions

pub fn memory_size(&self) -> usize

Memory size in bytes (4x reduction from f32)

Trait Implementations

impl Clone for QuantizedEmbedding

fn clone(&self) -> QuantizedEmbedding

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for QuantizedEmbedding

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

Formats the value using the given formatter.