What does a groundedness or faithfulness check measure in RAG?
Why A is the answer
Whether claims are supported by the retrieved context
It catches answers that go beyond the sources.
In detail
- It checks whether claims are supported by the retrieved context.
- It checks support, not truth in general.
Use case
An illustrative scenario, not a real incident.
A RAG answer adds a claim that is not in its sources, and a groundedness check flags it.
See it in different ways
flowchart TD
A{"What are you evaluating?"} -->|"Translation"| B["BLEU"]
A -->|"Summaries"| C["ROUGE"]
A -->|"RAG answers"| D["Faithfulness and relevance"]
A -->|"Open-ended quality"| E["Human or LLM judged rubric"]mindmap
root((Evaluation))
Translation
BLEU
Summaries
ROUGE
RAG
Faithfulness
Errors
Slice by topicflowchart TD
Q{"Is it in the retrieved context?"}
Q -->|"Yes"| A0["Grounded"]
Q -->|"No"| A1["Unsupported"]Key point: Groundedness checks support in context.
Why the other options are wrong
- B. Whether the answer is the longest of the candidates
- Length does not show support.
- C. Whether the response arrived within the latency target
- That is latency.
- D. Whether the question contains any spelling mistakes
- That is about the question.