Check Retrieval First When RAG Answers Go Wrong, Not the Prompt

A common reaction to a wrong answer from a documentation assistant is to tweak the prompt, but the model may never have received the correct page. A RAG system is a multi-stage pipeline (chunking, embedding, retrieval, reranking, generation), and errors can originate at any stage. The recommended approach is to log retrieved chunks for failing questions and determine whether the right information arrived at the model. Problems often trace to ingestion gaps, parsing losses, or chunks that lose context from their surrounding section. Developers should also check where the correct chunk ranks, since pure vector search can miss exact-match terms like error codes.