AHPQ.jl

Description

AHPQ.jl is a Julia library implementing the Anisotropic Hierarchical Vector Quantizer based on the Scalable Nearest Neighbours Algorithm (ScaNN). It follows the Google Research implementation (per the referenced paper and talk) and provides a configurable Max Inner Product Search (MIPS) system, written entirely in Julia with no non-Julia dependencies.

• Category: SDKs & Libraries
• Technology: Julia
• Use cases: Vector search, product quantization, MIPS-oriented indexing
• Source: https://github.com/AxelvL/AHPQ.jl

---

Features

• Anisotropic Hierarchical Vector Quantizer

  • Implements anisotropic hierarchical product/vector quantization.
  • Trains a vector quantization tree (VQ-tree) with an anisotropic loss function.
  • Optimized for Max Inner Product Search (MIPS) rather than standard L2 distance.

• ScaNN-based Algorithm

  • Based on Google Research’s Scalable Nearest Neighbours Algorithm (ScaNN).
  • Follows the approach described in the associated research paper and SlidesLive presentation.

• IVFADC-style Architecture

  • Can be viewed as an IVFADC setup (inverted file with asymmetric distance computation).
  • Uses vector quantization for coarse partitioning and refined distance evaluation.

• Julia-native Implementation

  • Written 100% in Julia.
  • No non-Julian dependencies, easing installation and integration into Julia projects.

• Configurable MIPS System

  • Provides a highly configurable Max Inner Product Search pipeline.
  • Designed for building high-performance vector search systems using product quantization.

(The repository content snippet does not expose further API-level or configuration details.)

---

Pricing

• Not applicable (open-source library; no pricing information provided in the available content).