QuestionQ38

VerifiedEvaluating and deploying agentic workflows

You are evaluating a multitool customer support agent built with Agent Development Kit (ADK) against a curated golden dataset of complex, multiturn user queries. During initial testing, the agent often returns plausible final answers but skips intermediate knowledge retrieval steps or produces responses unsupported by retrieved context, leading to ungrounded claims in production. You need to systematically isolate retrieval quality failures from response generation errors while preventing regressions in tool sequence trajectory accuracy.

What should you do?

Explanation

A Gen AI evaluation can assess whether responses are supported by retrieved context and compare tool-call trajectories with expected sequences, helping detect unsupported claims and skipped retrieval steps. (docs.cloud.google.com) Groundedness does not by itself measure retrieval completeness or relevance; separate retrieval-quality metrics are needed to fully isolate retrieval failures from generation errors.

Learn more

Community Discussion

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