A RAG app returns relevant passages, but answers add unsupported claims. Which evaluator targets this?
Why C is the answer
Groundedness
Groundedness measures whether the answer is supported by the retrieved context.
In detail
- Groundedness measures whether claims are supported by the sources.
- A fluent answer can still be invented.
Use case
An illustrative scenario, not a real incident.
Passages are right, but answers add unsupported claims. Groundedness scores this directly, where fluency and coherence would look fine.
See it in different ways
flowchart TD A["Foundry evaluators"] --> B["General: coherence, fluency"] A --> C["RAG: groundedness, relevance"] A --> D["Safety: hate, violence, protected material"] A --> E["Agents: tool call accuracy, task completion"]
mindmap
root((Evaluators))
Quality
Relevance
Coherence
Fluency
RAG
Groundedness
Agents
Tool call accuracy
Safety
Violence
Hate and unfairness
Protected materialsflowchart TD
Q{"What is the concern?"}
Q -->|"Unsupported claims"| A0["Groundedness"]
Q -->|"Reads badly"| A1["Fluency"]
Q -->|"Off the question"| A2["Relevance"]Key point: Unsupported claims are measured by groundedness.
Why the other options are wrong
- A. Fluency
- Fluency measures how well it reads.
- B. Coherence
- Coherence measures how well it hangs together.
- D. Violence
- Violence is a safety evaluator.