I wonder to what extent a non LLM system with access to the same original corpus of text would do, with some simple similarity search feature across the corpus?