Quantized vectors use 4x (int8) or 8x (int4) less memory than fp32. A randomized Hadamard rotation + scalar quantization packs vectors into smaller codes. An asymmetric query scores a full-precision query against quantized corpus in a WGSL kernel, then an exact fp32 re-rank recovers recall. This example is GPU-only — there is no CPU fallback for quantization.
ready.