For actual searching, it seems like RAG would be the way. Instead of rebuilding models, focus on curating datasets and sources.