next-plaid makes PLAID rescoring nearly free
Ex-AWS scientist posts that one int8 operation removes the slow rescoring step on existing indexes.
TLDR
Chao-Chun Hsu announced next-plaid changes that replace the exact rescoring stage in PLAID-style late-interaction search with a single int8 operation. The update runs on CPU, works with existing indexes, and removes the decompress segment shown in prior versions. A bar chart on the FiQA-15k set compares version 1.6.5 against the two new PRs at r=4. Omar Khattab retweeted the thread. No independent verification or performance numbers beyond the chart appear in the posts.
Combined views
7K
2 Sources, first seen 28d ago