You need a strict pass or fail check of whether RAG responses are consistent with the…

How it works

Foundry has two groundedness evaluators, and the difference is who does the judging:

• Groundedness gives a well-rounded definition that works with agent inputs, and uses your own GPT model as the LLM judge, which is why it takes a deployment_name. It returns Pass/Fail based on a threshold on a 1-5 scale.
• Groundedness Pro (preview) gives a strict definition, powered by Azure AI Content Safety and Microsoft's service model. It detects whether the response is strictly consistent with the context and returns True/False. Its inputs are query, response and context, with no deployment_name.

The requirements are strict pass/fail, consistency with the retrieved context, and no judge deployment of your own. That's Groundedness Pro.

The other options, and when you'd use them

• Retrieval: It judges how relevant the retrieved chunks are to the query, not whether the response is consistent with them, and it needs a judge deployment.
• Relevance: It measures how well the response addresses the query, not consistency with the context.
• Groundedness with your GPT deployment: It works, but it uses your own model as the judge, which the stem rules out.

Exam tip

Strict True/False groundedness with no judge deployment = Groundedness Pro (Content Safety). Your own GPT judge = Groundedness.

References

• Retrieval-Augmented Generation (RAG) evaluators