turbovec Beats FAISS IndexPQ on Recall at 14,151 Stars — TurboQuant's Data-Oblivious Rotation Compresses 10M Vectors From 31GB to 4GB

GitHub·medium signal

RyanCodrai/turbovec (MIT, Rust core with Python bindings) reached 14,151 stars implementing TurboQuant, a data-oblivious compression that applies a random rotation so every coordinate follows a known distribution regardless of input data, then does per-coordinate calibration, Lloyd-Max scalar quantization, and bit-packing for 16x compression at 1536 dimensions. Reported results: 0.2–1.9 points better R@1 than FAISS IndexPQ across 2-bit and 4-bit configs on OpenAI embeddings, 10–19% higher QPS than FAISS FastScan on ARM, ~5% faster on x86 4-bit but slower on x86 2-bit, and a 10M-document corpus shrinking from 31GB float32 to 4GB. Integrations exist for LangChain, LlamaIndex, Haystack, and Agno — the data-oblivious property matters because it removes the training/calibration step that makes PQ indexes brittle when the embedding distribution shifts.