RAG evals need misleading-context tests

A new RAG reliability paper evaluates behavior under clean, misleading, and mixed retrieval, including cases where retrieved context conflicts with a model's existing knowledge. Why it matters: Happy-path retrieval tests miss a common failure mode: irrelevant or poisoned context can increase confidence while pushing an answer away from the truth.

Try this: Add at least one mixed-context and one misleading-context case to a retrieval eval, then track whether the system cites, rejects, or over-trusts the bad context.