What's the difference between faithfulness, groundedness, and context presence?

Faithfulness asks whether the answer matches the source it cites, groundedness whether the answer came from the retrieved context at all, and context presence whether that context was retrieved in the first place.

They fail at different stages. Context presence fails in retrieval, before the model has written a word. Groundedness and faithfulness both fail in generation: an ungrounded answer came from the model’s own prior with the context sitting unused, which can still be correct whenever that prior happens to be right, and an unfaithful answer read the source and misrepresented it. None of the three says whether the retrieved source is itself accurate.

Knowing which one fired tells you where to look. A faithfulness failure means fix the generation step; a context-presence failure means fix retrieval.