TurboVec: Cost-Efficient Private Retrieval for Enterprise RAG

The paper presents TurboVec, an open-source vector index for enterprise RAG built on TurboQuant, a codebook-oblivious scalar quantizer that requires no corpus-dependent training. On the DBpedia OpenAI embeddings benchmark, using 1536-dimensional vectors across 100K–999K items, 4-bit TurboQuant improves Recall@5 by 8.5–8.9 percentage points over trained FAISS Product Quantization at the same memory budget. TurboVec reports 11 ms median latency at 100K vectors on Snowpark Container Services, versus 707 ms for a warehouse brute-force scan. Kernel-level allowlist filtering retains 0.86–0.93 Recall@10 across tenant workloads, while membership-inference accuracy is near random.