TurboQuant Pro compresses BGE-M3 with PCA

This is a strong reminder that embedding compression is often a basis-selection problem, not just a bit-budget problem. PCA is not a retrieval silver bullet, but it looks like a very solid linear baseline before moving to heavier codecs.

• –Naive tail truncation is a weak baseline for non-Matryoshka models; PCA makes the retained dimensions carry much more signal.
• –Cosine reconstruction alone is not enough for retrieval decisions, because Recall@10 degrades faster than similarity in the reported runs.
• –PCA plus low-bit quantization looks like a practical middle ground between scalar quantization and more aggressive binary or PQ compression.
• –The next questions are stability and generality: how much the PCA fit varies by corpus, language mix, and downstream ANN index behavior.