A knowledge-base team evaluates a grounded question-answering pipeline and wants a score that shows whether each claim in a generated reply is backed by the retrieved context. Which metric should they report?
Show answer and explanation
Answer: C
Groundedness evaluates whether statements in the response are supported by the context supplied to the model, which is the central concern for retrieval-augmented answers. Coherence judges how logically the text reads, while response length and time to first token are operational measurements. A reply can be fluent and coherent yet still invent facts, so faithfulness to the retrieved context needs its own metric.
Why the other options are wrong
- A. Length describes how long the output is. A short answer can still fabricate details, so length says nothing about whether content is supported by retrieved passages.
- B. Time to first token measures latency for streamed responses. It reflects performance, not whether generated claims match the source material.
- D. Coherence assesses how well sentences flow together. A fluent answer may still contain claims absent from the retrieved context, so coherence does not measure support.