TurboVec is an ultra-fast, lightweight vector indexing and similarity search library engineered in modern C++20 and Rust for maximum query throughput and deterministic low-latency vector similarity retrieval. Designed as a drop-in local vector storage engine for AI engineers, autonomous agents, and real-time inference pipelines, TurboVec eliminates the operational overhead of external client-server vector infrastructure by running directly inside the host process memory space.
The engine leverages hardware-native AVX-512, AVX2, and ARM NEON vector instructions to accelerate compute-intensive distance metrics, including Cosine similarity, Euclidean distance (L2), and Dot Product calculations. TurboVec implements a cache-conscious HNSW (Hierarchical Navigable Small World) index and an optimized Flat IVF index, supporting dynamic vector insertions, real-time deletion tombstones, and instant serialization to disk via memory-mapped binary storage.
The standout innovation of TurboVec is its native integration of Google's TurboQuant algorithm. In conventional vector quantization workflows (like Product Quantization), engineers must train a codebook over a large dataset sample. TurboQuant eliminates offline training through data-oblivious randomized rotations and scalar quantization, compressing 32-bit floats into 2-bit to 4-bit representations with 75% to 87.5% memory reduction while maintaining up to 98% retrieval recall.
TurboVec is 100% free and open-source under the permissive MIT license with over 16,400 GitHub stars. It provides ergonomic bindings for Python and Rust with arbitrary structured metadata filtering, making it an ideal local persistence layer for embedded RAG workflows, agentic short-term memory buffers, and desktop AI applications without external runtime dependencies.
