TL;DR
A high-performance vector index implementing the TurboQuant quantization algorithm in Rust, reducing memory usage to 1/8 while providing faster search speeds than FAISS.
Key features
TurboQuant-based quantization: Oblivious quantization independent of data distribution, enabling immediate indexing without separate training
High-speed SIMD kernels: Optimized for ARM NEON and x86 AVX-512BW, achieving 12-20% improvement over FAISS IndexPQFastScan
Filtered search: Limit search scope via allowlist or bitmask, with block-level early termination within SIMD kernels to minimize performance degradation
IdMapIndex: Supports addition/deletion/search based on external IDs, O(1) deletion
Purely local: No external services required, enabling a fully air-gapped RAG stack
When to use it
When you need to index large-scale embeddings (millions to tens of millions) memory-efficiently
When you need faster search than FAISS (especially on ARM environments)
When you need hybrid search (filtering then re-ranking) combined with external systems like SQL or BM25
For on-premises/VPC RAG applications where data privacy is critical