Bigger Context Windows Didn't Make Our RAG Smarter

We stopped measuring retrieval quality by how many tokens we could fit into the prompt. When long-context models became available, many of us made the same assumption. If an LLM can read 128K tokens, ...