RyanCodrai/turbovec

A vector index built on TurboQuant, written in Rust with Python bindings

What it solves

turbovec provides a high-performance, memory-efficient way to store and search massive collections of vector embeddings. It significantly reduces memory usage—storing vectors at 2 or 4 bits per dimension—while maintaining high search speeds and recall, making it suitable for local, in-process applications where large-scale vector storage is required without the overhead of a dedicated database server.

How it works

The library uses TurboQuant, a quantization technique that rotates vectors by a shared random orthogonal matrix to ensure coordinates follow a predictable distribution. It then quantizes these coordinates using a pre-computed Lloyd-Max codebook and packs them into bits. At search time, it uses SIMD-optimized kernels to perform fast inner-product searches directly on the compressed data. It supports incremental updates, atomic file synchronization, and filtering via allowlists or bitmasks.

Who it’s for

Developers building local AI applications in Python or Rust who need to perform fast similarity searches on large datasets (e.g., millions of embeddings) but are constrained by RAM. It is designed for those who prefer an in-process library over a standalone vector database service and are comfortable with a lossy, quantized search approach.

Highlights

  • Extreme Compression: Reduces memory footprint by up to 15.5× compared to float32 storage.
  • High Performance: Searches are significantly faster than FAISS FastScan on both ARM and x86 architectures.
  • Zero-Training: Uses a mathematical codebook, meaning no training phase or index rebuilding is required when adding new data.
  • Drop-in Integrations: Compatible with popular frameworks like LangChain, LlamaIndex, Haystack, and Agno.
  • Local & Durable: Operates entirely within your process and supports atomic, crash-safe persistence.

Related

  • Dispatch
    Turbovec: High-Performance Vector Search using Google's TurboQuant

    Turbovec is a Rust-based vector index implementing the TurboQuant algorithm, offering up to 16x memory compression and faster search speeds than FAISS IndexPQFastScan.

  • Project
    timescale/pg_vectorsearch

    A PostgreSQL extension providing a high-performance, scalable vector search index (PRISM) that uses hierarchical clustering and quantization to outperform standard HNSW and IVFFlat methods.

  • Project
    SeekStorm/SeekStorm

    A high-performance Rust-based search engine providing sub-millisecond native vector and lexical search, available as an in-process library or a multi-tenancy server.

  • Project
    facebookresearch/faiss

    A library for efficient similarity search and clustering of dense vectors, capable of handling billion-scale datasets using CPU and GPU acceleration.

  • Project
    HKUDS/LightRAG

    LightRAG is an open‑source Python framework that blends vector embeddings with a knowledge‑graph index for fast, scalable Retrieval‑Augmented Generation. It supports multiple chunking strategies, multimodal document parsing, incremental updates, and a variety of storage back‑ends, and ships with a REST API and Web UI.