A technique that feeds the model sources at query time so it answers from trusted material and hallucinates less.