Technology method · wiki:deep

Embedding quantization

Reduce bits per dimension (int8, binary) — complementary to Matryoshka length truncation; mxbai documents both.

Embedding quantization flow

Quantization changes the numeric type of each coordinate. The vector length d usually stays (unless the product also truncates). That is why OpenAI’s dimensions=1024 and mxbai’s dimensions=512 are not quantization: they change how many floats you keep.

Operators often stack both: Matryoshka to a smaller m, then int8/binary for RAM. FAISS and service vector DBs expose product-quantization / scalar quantization paths under ANN indexes.

See also

In this taxonomy

_catalog/taxonomy/tech-embedding-quantization.md