A 311 engineering team sees a cluster of poor answer classes and needs to inspect stored prompt and response fields. Which troubleshooting tooling approach helps?

When a 311 engineering team sees a cluster of poor answer classes, it needs to inspect stored prompt and response fields. Querying CloudWatch Logs Insights over logged prompts and responses can isolate patterns such as bad retrieval chunks, weak instructions, or recurring refusal and edge cases, without resorting to blunt infrastructure actions. Structured log analysis is the appropriate troubleshooting approach for GenAI quality regressions in civic assistants. Deleting all GenAI logs immediately fails because it destroys evidence needed for diagnosis and may violate retention needs for service improvement, subject to privacy policy. Treating the cluster as transient and restarting every service in the Region fails because regional restarts do not fix systematic prompt or response defects, and this approach skips review of the stored content. Asking residents to mail printed screenshots without structured log analysis fails on scalability and makes it difficult to aggregate evidence across a poor answer class. Exam caveat: prompt and response logging must follow municipal privacy and retention rules; redact PII where required, restrict access, and avoid unbounded public exposure. Scope Logs Insights queries by time, answer-class tags, and request IDs to avoid over-collection. Logging is not a substitute for automated evaluation suites. Practical check: tag a sample of poor 311 answers, query prompts, retrieved context, and responses for correlations, identify a repeating failure pattern, apply a prompt or retrieval fix, and query again to confirm the poor answer class shrinks without deleting the audit trail.