The arrangement depends on the presence of estimators or transformers in the pipeline. For instance, with estimators like StringIndexer, refitting is necessary each time if the pipeline is placed inside the cross-validator. However, to prevent data leakage from initial steps, it's safer to position the pipeline within the cross-validator. This approach ensures the cross-validator first splits the data before fitting the pipeline, avoiding potential information leakage from the hold-out set to the training set.