Matryoshka truncation
An embedding drawn as a row of dimensions, darkest at the front because a Matryoshka model stores the most important information in the first dimensions. Truncating keeps a prefix of the vector. The kept head is what stays in RAM for the first retrieval stage, while the full vector stays on disk.
DIMENSIONS
KEPT
RAM
Disk
256
of 1536 dims
17
% of full size
Keep