Struct Reranker 

pub struct Reranker<'a> { /* private fields */ }

Reranker for computing int8 dot products with outlier correction

Implementations

impl<'a> Reranker<'a>

pub fn new( i8_data: &'a [i8], scales: &'a [f32], outliers: &'a [OutlierEntry], dim: usize, num_outliers: usize, ) -> Self

Create a new reranker

pub fn score(&self, vid: VectorId, query_i8: &[i8], query_scale: f32) -> f32

Compute dot product score for a single candidate

Uses SIMD-accelerated C++ kernels via FFI when available:

• AVX2: 32 int8 ops per cycle (8x speedup for dim=768)
• AVX512: 64 int8 ops per cycle (16x speedup)
• NEON: 16 int8 ops per cycle (4x speedup)

pub fn score_with_fp32( &self, vid: VectorId, query_i8: &[i8], query_scale: f32, query_fp32: Option<&[f32]>, ) -> f32

Compute dot product score with optional fp32 query for accurate outlier computation.

When query_fp32 is provided, outlier contributions use exact fp32 values instead of reconstructing from quantized int8, reducing error from O(1/127) to floating-point epsilon.

Arguments

• vid - Vector ID to score
• query_i8 - Quantized query vector (for main dot product)
• query_scale - Query quantization scale
• query_fp32 - Optional original fp32 query (for accurate outlier scoring)

pub fn score_batch( &self, candidates: &[VectorId], query_i8: &[i8], query_scale: f32, ) -> Vec<ScoredCandidate>

Score multiple candidates in batch

pub fn score_batch_with_fp32( &self, candidates: &[VectorId], query_i8: &[i8], query_scale: f32, query_fp32: &[f32], ) -> Vec<ScoredCandidate>

Score multiple candidates with fp32 query for accurate outlier computation

pub fn rerank( &self, candidates: &[VectorId], query_i8: &[i8], query_scale: f32, r: usize, ) -> Vec<ScoredCandidate>

Rerank and return top R candidates

pub fn rerank_with_fp32( &self, candidates: &[VectorId], query_i8: &[i8], query_scale: f32, query_fp32: &[f32], r: usize, ) -> Vec<ScoredCandidate>

Rerank with fp32 query for accurate outlier computation