One of the most expensive debugging mistakes in AI systems is assuming the model is the problem. A user receives a bad answer. The response looks wrong. The immediate reaction is usually: "The model hallucinated." Sometimes that is true. Many times it is not. One production incident reminded us of that very clearly. What initially looked like a model quality issue turned out to be a retrieval problem hiding underneath. Everything Pointed at the Model The first reports were straightforward. Users said the system was giving incomplete answers. Not completely wrong. Just missing important information. At first glance, it looked like a reasoning problem. The responses were: shorter than expected missing key details inconsistent across similar questions Nothing crashed. No errors appeared. Latency remained normal. Infrastructure metrics looked healthy. The obvious suspect was the model. Prompt Testing Didn't Change Anything The first thing we tried was what many teams would try. Prompt investigation.…