Turbovec – A Rust-based Vector Search Engine for Google TurboQuant

turbovec, a Rust-based vector index built on Google Research's TurboQuant algorithm, has been released. This tool is optimized for developers who prioritize memory usage, latency, and security when building RAG (Retrieval-Augmented Generation) stacks.

Key Features:

• Overwhelming Memory Efficiency: Compresses a 10-million-document corpus from 31GB (float32) to 4GB, providing faster search speeds than FAISS.
• Data-oblivious Quantizer: Maintains optimal distortion rates without separate training steps or parameter tuning, enabling Online ingest without retraining or rebuilding as data grows.
• SIMD-optimized Search: Demonstrates an average 3.4x (4-bit) performance improvement over FAISS IndexPQFastScan through hand-written kernels for ARM (NEON) and x86 (AVX-512, AVX2) architectures.
• Functional Flexibility: Supports Filter at search time via ID whitelists and ensures crash safety and efficiency with Incremental saves, which store only changed parts.
• Local and Security-focused: Operates locally or within a VPC without managed services, making it suitable for building air-gapped RAG stacks.

Framework Integration: Supports drop-in replacement compatible with vector stores in major AI frameworks such as LangChain, LlamaIndex, Haystack, and Agno, allowing easy migration of existing pipelines.