Why RAG and how we will judge it

Learn the idea: the LLM does not know your company documents, and retraining is costly. RAG = first search your documents for the relevant parts (retrieval — you built this in the NLP module's FAQ search), then put those parts in the prompt and let the LLM answer from them (generation). Scores: does retrieval find the right page, is the answer correct, is every statement supported by the sources, does it refuse when the answer is not in the documents.