What do you predict when shipping a generative retrieval prototype from a 200k-document set to a 20M-document corpus?

As the corpus grows, the bits available per document decrease. At 20M documents, the model has significantly fewer bits per document compared to the 200k set, likely falling below the threshold required for effective retrieval. Furthermore, identifier reassignment means the original 200k results cannot be used as a baseline. A controlled experiment sampling the corpus is necessary to understand the degradation curve.