[nblgbg]

I agree with the comment. Especially if you index these quantized vectors with HNSW you will get even less precision. You need to retrieve lot more and do a kNN with the FP32 vectors is the way to go !