$ retrieval-augmented generation

RAG, measured against ground truth

Ask a question with and without a corpus. The model never changes — only what it is allowed to read. Where the corpus ships a known-correct answer, every question is scored: you can see exactly which ones the retriever gets right and which it misses.