Embedding Vector Oriented Clustering (EVōC) is a Python library designed for rapid and flexible clustering of large, high-dimensional embedding vectors. It targets users working with embeddings from models like CLIP, sentence-transformers, OpenAI, and Cohere, offering significant speed improvements and reduced hyperparameter tuning compared to traditional methods like UMAP + HDBSCAN. The library excels at producing high-quality clusters efficiently, even for quantized vector formats.

How It Works

EVōC specializes in embedding vectors, optimizing its approach to bypass the time-consuming aspects of general-purpose clustering pipelines. It leverages techniques inspired by PLSCAN and density-based methods to achieve fast, CPU-bound clustering. A key innovation is its ability to generate multi-granularity clusters, providing a hierarchy of results from fine-grained to coarse-grained, and it natively supports clustering of int8 or binary quantized embeddings.

Quick Start & Requirements

Installation is straightforward via pip: pip install evoc. Core dependencies include numpy, scikit-learn, numba, tqdm, and tbb. No specialized hardware like GPUs is mentioned. Full documentation is available at https://evoc.readthedocs.io/en/latest/.

Highlighted Details

• Achieves fast clustering of embedding vectors on CPU.
• Supports multi-granularity clustering with automatic cluster number selection and hierarchy extraction.
• Natively handles clustering of int8 or binary quantized embedding vectors.
• Includes automatic detection of duplicate or near-duplicate vectors.

Maintenance & Community

The project welcomes contributions via pull requests. Specific community channels (e.g., Discord, Slack), roadmap, or notable contributors/sponsorships are not detailed in the README.

Licensing & Compatibility

EVōC is released under the permissive BSD (2-clause) license, allowing for broad compatibility, including commercial use.

Limitations & Caveats

The library is explicitly described as an "early beta version," with a warning that "Things can and will break right now." Users should expect potential instability and are encouraged to provide feedback.