turbovec is a production-ready vector search index written in Rust with Python bindings, built on Google's TurboQuant quantization. The pitch: compress 10M documents from 31 GB (float32) to 4 GB while searching faster than FAISS. It offers no-training online insertion, SIMD search kernels hand-tuned for ARM and x86, incremental persistence (one fsync per write, milliseconds even at large scale), and runtime filtering without recall loss. The core claim rests on TurboQuant's data-oblivious quantization—after random rotation, coordinates follow predictable distributions regardless of input, so optimal Lloyd-Max codebooks are precomputed rather than trained. Optional calibration (TQ+) fits per-coordinate shift/scale to empirical quantiles for 0–2pp recall gains on low-bit settings. Benchmarks show 3.4× faster search than FAISS IndexPQFastScan at 4-bit across two CPU architectures, 20–26% gains at 2-bit, and dramatically faster insertion (7–14× on single vectors). The source includes framework integrations (LangChain, LlamaIndex, Haystack) and cites the ICLR 2026 TurboQuant paper. One open question: how sensitive are those speed wins to workload variance—do the hand-written kernels maintain their edge on non-OpenAI embeddings and atypical access patterns in production RAG systems?