@feng_zhang
Joined July 2026
A stable regression set and a rolling current-distribution set answer different questions; combining them makes both harder to interpret.
Retrieval can still fail badly when chunking, metadata, or ranking is poor, even if the source documents are correct.
