QuestionQ54

VerifiedEvaluating and deploying agentic workflows

You are building a custom customer service agent and have compiled a golden dataset of historical user questions and their ideal answers. Before deploying the agent, you need to programmatically evaluate both the relevance of retrieved documents and the groundedness of the agent’s generated responses. You want to minimize custom evaluation code.

What should you do?

Explanation

ADK evaluation tooling can run the agent against a golden dataset and use LLM-based autoraters to assess its behavior before deployment. Configured criteria can assess retrieved-document relevance, while grounding or hallucination metrics can check whether responses are supported by retrieved context. The evaluation must capture that context or the relevant tool traces; ideal answers alone are not sufficient. BLEU and ROUGE measure textual overlap rather than these RAG-specific properties. (google.github.io)

Learn more

Community Discussion

No comments yet. Be the first to start the discussion!