Module vector 

Vector indexing data structures

This module provides a modular architecture for vector search:

Module Structure

• ivf - Core IVF (Inverted File Index) infrastructure

  • CoarseCentroids - k-means clustering for coarse quantization
  • SoarConfig / MultiAssignment - SOAR geometry-aware assignment

• quantization - TurboQuant training-free codec

  • TqCodec - derived rotation/codebook, no trained artifacts

• index - Segment payloads for the production ANN implementations

  • IvfTqIndex - float vectors with TQ-coded centroid residuals
  • BinaryIvfIndex - exact packed binary vectors

SOAR (Spilling with Orthogonality-Amplified Residuals)

The IVF module includes Google’s SOAR algorithm for improved recall:

• Assigns vectors to multiple clusters (primary + secondary)
• Secondary clusters chosen to have orthogonal residuals
• Improves recall by 5-15% with ~1.3-2x storage overhead

Re-exports

pub use ivf::CoarseCentroids;

pub use ivf::CoarseConfig;

pub use ivf::IvfProbePlan;

pub use ivf::MultiAssignment;

pub use ivf::SoarConfig;

pub use quantization::TqFlatBuilder;

pub use quantization::TqCodec;

pub use quantization::TqQueryPlan;

pub use index::BinaryCoarseQuantizer;

pub use index::BinaryIvfConfig;

pub use index::BinaryIvfIndex;

pub use index::IvfTqIndex;

pub use index::TqIvfEncodeScratch;

pub use index::TqIvfQueryPlan;

pub use index::is_ivf_tq_cosine_generation;

pub use index::mark_ivf_tq_cosine_generation;

Modules

index

Vector index implementations

ivf

IVF (Inverted File Index) module for vector search

quantization

Vector quantization for IVF indexes