Combine IVF clustering with TurboQuant int8/int4 quantization. This is the 1M-vector path: quantization shrinks each vector 4-8x, while IVF scans only a fraction of clusters. Together they make 1M x 768 fit in ~1 GB and query in under 10 ms. This example is GPU-only — neither quantization nor IVF has a CPU fallback.
ready.