made a small library using GPT5.5-Pro and autoresearchyou can convert 384-dim f32 vectors go from 1536 bytes to 48 bytes without calibration. works for petabyte scale processing of text in pure online manner.