Given eighteen months of anonymized query logs, should...

Real query logs are superior to proxy-based pseudo-instructions, but they are limited by availability. Logs are corpus-level, while anchors are document-level, so an attribution step (like clicks or top-1 retrieval) is necessary. For new documents without logs, pseudo-instructions provide the necessary structure. It is important to mitigate feedback loops where the system only chunks based on what the current retriever already surfaces by holding out a slice of data for evaluation.