TURBOQUANT: Data-oblivious vector quantization benchmarks PolarQuant and QJL for biomedical embedding compression

Biomedical RAG systems need compressed embedding indices. TURBOQUANT benchmarks PolarQuant (random rotation + scalar quantization) and QJL (1-bit Johnson-Lindenstrauss projection) on synthetic 256-dim embeddings. Results: 4-bit = Recall@10 0.814, cosine 0.990, 8x compression; 3-bit = 0.628, 10.6x; 2-bit = 0.364, 15.9x. Limitation: synthetic data, random rotation not corpus-optimized, brute-force search. Ref: Chen J et al. TurboQuant arXiv:2504.19874 (2025); Johnson-Lindenstrauss Contemp Math 1984. Authors: Zamora-Tehozol EA, DNAI.