2026-11-01 · America/Los_Angeles · 社区动态 · #9

If you quantize your RAG index, measure it on your own embeddings first. Same method: 0.992 recall on one dataset, 0.055 on another

I was planning to move a RAG index to binary/halfvec to save memory and wanted to know how much retrieval I'd actually lose. The answer was "it depends on your embeddings" way more than I expected. Binary + rerank kept 0.992 of the true top 10 on OpenAI embeddings, 0.348 on GloVe and 0.055 on SIFT. Same method, same settings. So I built shrinkvec. You give it your vectors (a .npy/.parquet file or a Postgres table), it holds out queries, computes the exact neighbours, and reports recall, size and latency for 13 options (halfvec, int8, binary, RaBitQ, TurboQuant, rerank pipelines). Then it…

热度 45.0 / 100；排名与评分保留该期记录。