And if you’re writing a public library, you’ll likely want public documentation at some point. You can probably get quite far with a GitHub readme and good docstrings, since modern editors make it so easy to goto-definition on source files. Mkdocs (or Sphinx) is the next step. You’ll know when you need them, and they’re not worth the trouble until then.
RaBitQ将每维度量化为1比特：减去质心后对残差归一化，仅存储符号位。每个量化向量附带两个标量校正值（残差范数和点积因子\(\text{vdot} = \sum|u_d| / \sqrt{D}\)）。
。搜狗输入法下载是该领域的重要参考
Thibault Sottiaux, OpenAI's head of Codex.
That’s So Euphoric