Enterprise AI systems that answer questions from documents depend on two things working: retrieval and generation.

Most evaluation approaches measure the end result. They score whether the answer was right or wrong. When the answer is wrong, the diagnosis is usually one of two defaults: the model is not good enough, or the retrieval is not good enough.

This study asks a different question: When a grounded AI system fails, how often is the failure in retrieval, and how often is it in utilization?

Download the latest research from Seekr to learn how often does the system retrieve the correct source document and still produce the wrong answer.