Qinco

Brand: Meta AI Research
Category: SDKs & Libraries
Source: GitHub – facebookresearch/Qinco

Qinco (and its improved variant QINCo2) is an open‑source, neurally‑augmented residual quantization system for compact vector representations. It is designed as a low-level component for vector indexing, compression, and large‑scale similarity / nearest neighbor search.

---

Overview

Qinco implements Quantization with Implicit Neural Codebooks (QINCo) and its successor QINCo2: Vector Compression and Search with Improved Implicit Neural Codebooks. It targets multi‑codebook vector quantization—specifically residual quantization (RQ)—to compress high‑dimensional vectors while maintaining accuracy for similarity and nearest neighbor retrieval at scale.

The repository includes:

• QINCo2 (latest implementation and improvements)
• QINCo1 (original ICML 2024 implementation, under qinco_v1)

---

Features

Core Quantization Approach

• Neurally-augmented multi‑codebook vector quantization
  • Implements residual quantization (RQ) with learned implicit neural codebooks.
  • Replaces fixed codebooks at each quantization step with a neural network that predicts the next codebook conditioned on the already‑quantized prefix of the vector.
• Implicit neural codebooks conditioned on Voronoi cells
  • Codebooks depend on previously selected Voronoi cells, increasing the expressive capacity of the compression without storing large explicit codebooks.
• Dynamic rate quantization
  • Each quantization step is trained with its own quantization error objective.
  • A trained system for a given compression rate can also be used at lower compression rates, enabling flexible trade‑offs between compression and accuracy.

QINCo2 Improvements over QINCo1

• Fast approximate encoding
  • New approximate encoding procedure.
  • Achieves similar mean squared error (MSE) as the original method.
  • Significantly reduces training and encoding time.
• Beam search integration in encoding
  • Adds beam search to the encoding pipeline.
  • When combined with approximate encoding, achieves much lower compression error than QINCo1 at similar encoding time.
• Optional pairwise decoder module for retrieval
  • New (optional) module in the large‑scale retrieval pipeline.
  • Uses a pairwise decoder to improve retrieval accuracy.
• Architecture and training pipeline upgrades
  • Overall improvements to the model architecture.
  • Refined training process for better compression and search performance.

Large-Scale Retrieval & Search

• Designed for similarity and nearest neighbor search
  • Produces compact vector representations suitable for approximate nearest neighbor (ANN) search.
  • Targets large-scale datasets and vector databases / AI retrieval systems.
• Integration into retrieval pipelines
  • Supports large‑scale vector indexing and retrieval scenarios.
  • Optional pairwise decoder can be added to improve search quality.

Data & Experiment Support

• Reference datasets and scripts
  • Download scripts for common large-scale benchmarks:
    • BigANN (with -small option for reduced storage use)
    • Deep1B (with -small option)
    • Contriever
    • FB-ssnpp
• Pretrained checkpoints (base experiments)
  • Checkpoints provided for QINCo1 and QINCo2 base experiments (as referenced in the repository).

Implementation & Setup

• Open-source Python implementation
  • Requires Python 3.
  • Dependencies managed via environment.yml / requirements.txt.
• Conda-based environment setup
  • Example workflow:
    • git clone https://github.com/facebookresearch/Qinco (or updated repo path as in docs)
    • cd Qinco
    • conda env create -f environment.yml

Research Orientation

• Reproducible research code
  • Includes code to reproduce results from:
    • ICML 2024 paper: Residual Quantization with Implicit Neural Codebooks (QINCo1).
    • ICLR 2025 paper: QINCo2: Vector Compression and Search with Improved Implicit Neural Codebooks.
• Citable resource
  • Provided BibTeX entry for citing QINCo2 in research.

---

Use Cases

• As a low-level quantization and indexing component for:
  • Vector databases.
  • Large-scale AI retrieval and recommendation systems.
  • Embedding-based search in NLP, CV, and multimodal applications.
• For research and experimentation on:
  • Neural vector quantization methods.
  • Compression–accuracy trade‑offs in large‑scale ANN search.

---

Pricing

• Open-source: No pricing information is listed; the project is distributed as open-source research code on GitHub.

---

Tags

• vector-compression
• similarity-search
• open-source