Retrieval & RAGintermediate12 min · 9 steps
Answer Generation
The last step of RAG: generate an answer that stays inside the evidence, cites it claim by claim, follows a template or JSON schema, reasons privately, and synthesises across documents.
claims supported72%
parseable81%
requesterrorresponsesignaldatatoken
metrics are illustrativeProblemStep 1 of 9
Good evidence, bad answer
Retrieval found the right chunks, and the answer still fails: it blends the evidence with the model's general knowledge, cites nothing, rambles in a different format every time, and breaks the JSON the ticketing system expects.
Answer generation is where retrieved context becomes a response — and it needs its own rules.
Lesson 31 of 37 in Retrieval & RAG