How do I tell a stale RAG index apart from a truncated context window?

Look at faithfulness, not just the final answer. A stale index passes faithfulness and fails groundedness: the model answers exactly what the retrieved chunk says, so it’s consistent with its own context, but the chunk itself describes a policy or price that changed since the index was last built, so the claim doesn’t hold up against reality. A truncated context window fails faithfulness itself: the relevant chunk was retrieved correctly, but got cut off before the model saw it, or crowded out by other chunks competing for the same window, so the answer no longer traces back to anything it actually received.

The practical test is whether the retrieved chunk, read on its own, still supports the answer. If it does and the answer is wrong, look at when the index was last built. If it doesn’t, look at what got cut before generation.