Definition
Plain language
Sticking to the documents you were handed instead of what you remembered from training.
As stated in the literature
The degree to which a model's answer follows the retrieved context when that context conflicts with its parametric knowledge; a design goal for RAG that is difficult to distinguish from susceptibility to poisoned evidence.
Also called: context faithfulness, context-faithful
Why it matters: It is the whole point of feeding a model fresh documents — but the same obedience means a single planted or wrong document can talk the model out of a correct answer.
For example, if the documents handed to a model say a company's headquarters moved to a new city, a context-faithful answer reports the new city instead of the one the model memorized during training.
Heard on the show
“And there's a whole research line — context-faithfulness — that treats exactly that behavior as the goal, because RAG exists to override stale or wrong parametric knowledge.”Episode 247 — One Edited Photo, an Honest Caption, and a RAG System That Believes It