Start with a question set
Collect representative questions from documentation searches, onboarding and support. Include ambiguous requests, version-specific details and questions the available sources cannot answer.
Score evidence before prose
Check whether the retrieved passages contain enough information to answer. A fluent response cannot repair missing evidence. Track source coverage separately from answer style.
Review the failure modes
Distinguish missing content, outdated sources, poor retrieval and unsupported generation. Each failure needs a different fix; use the findings to decide which part of the workflow to improve.