Fidelity, also called groundedness here, measures whether generated answers are supported by the retrieved context. In a RAG chatbot, it helps assess whether responses stay anchored to information they were expected to use. section 1.3.4 section 2.3.6
Compare model versions using a fixed evaluation setup and a defined metric. An LLM judge can score groundedness, but its output is an evaluation signal, not absolute truth; validate and interpret it carefully. section 1.3.4 section 1.4
Make results useful to product decisions by translating a score change into plain language—for example, more generated answers were supported by retrieved context. A benchmark improvement alone may not capture live product behavior. section 2.3.6 section 1.4










