My results suggest that the main reason their entire thing works is because they compare substrings between some query text and all texts in the corpus. By precompiling the list of substrings that occur in the corpus, I can get their algorithm to have linear complexity at inference time, instead of quadratic complexity as described in the paper.