Open Questions and Risks

Questions

• Which GCP project, region, quotas, and billing controls are used in practice?
• Is Firebase auth authoritative in production or is JWT auth still primary?
• What are the supported LLM providers and fallback behavior?
• How are notifications and escalation handled after the PoC?
• Are Cloud Run and GCS audit logs retained and reviewed?
• What are the target SLA, latency, and RAG relevance metrics?

Risks and next steps

Risk	Next step
Placeholder secrets	Fail closed when production secrets are missing
RAG access regression	Add cross-role authorization tests
LLM extraction errors	Validate structured outputs and add repair/retry policy
Upload abuse	Enforce size, MIME, malware, and parser time limits
Cloud cost drift	Add budgets, alerts, and Cloud Run/SQL usage dashboards
PoC operational gaps	Define SLOs, backups, restore drills, and incident runbooks

Verification checklist

• Confirm production authentication authority and token lifecycle.
• Exercise a complete citizen-to-officer case in a deployed environment.
• Verify GCS objects are private and downloads require authorization.
• Run cross-role RAG tests with intentionally sensitive fixtures.
• Test Cloud Run job retries and partial deployment recovery.
• Measure chat first-token latency, finalization latency, and RAG relevance.
• Establish database backups, restore validation, and retention windows.
• Add notification and SLA escalation design before production launch.