Given a 143 GiB ColBERT index and a strict budget, how...

When optimizing a large index, you must approach the problem as a trade-off between storage and retrieval quality. By calculating the cost-per-point of MRR@10, you can prioritize cuts that yield the best performance-to-size ratio. For instance, dimension reduction offers a much better return on investment than token collapse. It is also important to be prepared to defend why you chose not to use pooling, which often incurs a much higher performance penalty than the chosen dimension reduction strategy.