Struct AdaptiveQuantizer 

pub struct AdaptiveQuantizer { /* private fields */ }

Adaptive quantizer that supports per-vector bit-width selection and tier promotion/demotion.

Wraps three TurboQuantizer instances (hot/warm/cold) and provides methods to encode at a specific tier, promote to higher precision, or demote to lower precision.

Implementations

impl AdaptiveQuantizer

pub fn builder(dim: usize, seed: u64) -> AdaptiveBuilder

Create a builder for an adaptive quantizer.

Arguments

• dim — Vector dimension
• seed — Random seed for deterministic quantization

pub fn encode_adaptive( &self, vector: &[f32], tier: Tier, ) -> Result<AdaptiveCode>

Encode a vector at the specified tier.

pub fn decode_adaptive(&self, code: &AdaptiveCode) -> Vec<f32>

Decode an adaptive code back to an approximate vector.

pub fn inner_product_estimate( &self, code: &AdaptiveCode, query: &[f32], ) -> Result<f32>

Estimate inner product between a code and a query vector.

pub fn promote(&self, code: &AdaptiveCode) -> Result<AdaptiveCode>

Promote a code to a higher-precision tier.

Decodes the current code and re-encodes at the next tier up. Hot codes cannot be promoted further.

Note: Promotion operates on the approximate reconstruction, not the original vector. The result is lower quality than encoding the original vector directly at the target tier. Use when the original is unavailable.

pub fn demote(&self, code: &AdaptiveCode) -> Result<AdaptiveCode>

Demote a code to a lower-precision tier.

Decodes the current code and re-encodes at the next tier down. Cold codes cannot be demoted further.

pub fn dim(&self) -> usize

Return the vector dimension.