Struct Turbovec 

pub struct Turbovec { /* private fields */ }

turbovec 2/3/4-bit TurboQuant encoder.

The on-row payload is the original vector’s f32 little-endian bytes. Compression and SIMD scoring happen at the per-table layer (see [crate::index::TurboTable]) where a turbovec::TurboQuantIndex holds every vector in its 2/3/4-bit packed form. Per-row storage stays at 4 * dim bytes; the in-memory ANN index achieves the headline 8x (4-bit) or 16x (2-bit) compression on the slot table.

The on-row passthrough is deliberate: turbovec’s public API does not expose per-vector reconstruction, so a faithful decode implementation requires keeping the source bytes. The trade-off (no row-layer compression) is documented in docs/dynvecdb/architecture.md.

Implementations

impl Turbovec

pub fn new(bits: u8) -> Self

Build a turbovec encoder for the given bit width.

Panics

Panics on a bit width outside {2, 3, 4}. Use Codec::encoder in the dynamic-dispatch path to keep the panic out of caller code; the codec variants are pre-validated.

pub fn bits(self) -> u8

Bit width this encoder was built with.

Trait Implementations

impl Clone for Turbovec

fn clone(&self) -> Turbovec

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Turbovec

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

Formats the value using the given formatter.

impl Encoder for Turbovec

fn codec(&self) -> Codec

Codec identifier produced by this encoder.

fn encode(&self, values: &[f32]) -> Result<EncodedVector, EncodingError>

Encode values into an EncodedVector.

fn decode(&self, ev: &EncodedVector) -> Result<Vec<f32>, EncodingError>

Decode an EncodedVector back to Vec<f32>.

impl Copy for Turbovec