I have another project where I was doing semantic search via standard RAG and realized an agent plus grep was actually way better than matching user text to obscure legalese vectors. Claude already knows CFR from its training data so targeted grep recalls data with less ranking gymnastics.