Groundedness

Check if responses are grounded in provided context.

The Groundedness detector uses an LLM to verify that assistant responses are factually grounded in the provided context. Use it to catch hallucinations in RAG applications.

Recommended for Output

Use cases

  • Detect hallucinations in RAG-based applications
  • Ensure responses only cite provided documentation
  • Verify factual accuracy before showing to users

Labels

NOT_GROUNDED

The response contains claims not supported by the provided context.

Configuration

Threshold default: 4

Minimum grounding score (0-5). Responses scoring below this are flagged.

The reference context must be passed via the API request metadata.