[kordlessagain]

Have an LLM generate a reverse index for the entries, but force it to keep the cardinality low. Then you can use a Jaccard similarity.