It's so important to look at the actual prompts sent to the model (and the received chunks) to see if it makes sense at all. A lot of RAG pipelines retrieve "relevant" chunks that are actually useless for the specific question. The routing agent is where most of the value is IMO.