I was working at a startup that was developing an annotation pipeline for text and I thought a lot about what set of operators you'd use to do things like: modify the text with index-controlled labels and keep the labels aligned, change the tokenization of labeled data (labeled per token) w/o having to re-label the data, etc.