Compress vector indexes with int8 (4×) or binary quantization (32×) before scaling out

FreeAcademy·medium signal

Scalar quantization to int8 yields ~4× index compression for roughly 1% recall loss; binary quantization shrinks indexes 32×. Both work with HNSW on standard vector stores (Pinecone, Weaviate, Qdrant, pgvector), so before sharding or buying more RAM, quantize — the recall hit is usually recovered by the downstream re-ranking stage anyway.