179 questions
No questions match those filters.
← All questions
RAGMedium· Retrieval-Induced Harm· Model Knowledge
Name a class of questions where turning retrieval on ma...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansRetrieval’s benefit can be modeled as a rescue term (when the model doesn’t know the answer) and a harm term (when the model is misled by retrieved content). As the probability of the model already knowing the answer approaches 1, the upside of retrieval vanishes, but the risk of being misled remains. This results in roughly 10% of questions being flipped from correct to wrong.