When we use all EDAM classes in the embedding space, during the retrieval of the top-K classes, we may have an imbalanced proportion of Data / Formats / Operation / Topics. The idea here is to decompose the process and search only for EDAM operations.