Review Packet: GenUI and AI System
When to use
Use this packet for generative UI, AI copilots, RAG features, agentic workflows, tool calling, MCP/A2A integrations, AI-generated components, and model-driven workflow automation.
Intake questions
- Which parts of the experience are deterministic and which are AI-assisted?
- What component registry can the model select from?
- What schema validation happens before rendering?
- Which retrieval sources are allowed and how are citations shown?
- Which tools can be called and which require approval?
- What happens when the model is wrong, slow, unavailable, or unsafe?
- Which evals block release?
Required artifacts
- GenUI component registry contract.
- Streaming event protocol.
- RAG source and citation policy.
- Tool risk matrix and approval flow.
- Eval suite and red-team cases.
- AI observability scorecard.
Decision matrix
| Need | Architecture control |
|---|---|
| Generated UI | Registry-only components with schema validation. |
| Knowledge answers | RAG with citations, freshness rules, and source quality checks. |
| User actions | Risk-classified tools with permission propagation. |
| Irreversible actions | Human approval and audit log. |
| Model uncertainty | Fallback UI, escalation, and user correction capture. |
Red flags
- Model can invent arbitrary component names or props.
- Tool execution depends on frontend-only permission checks.
- Citations are absent for policy or compliance answers.
- Eval failures are tracked but do not block release.
- Prompt injection is treated as a prompt wording problem only.
Approval criteria
- AI boundaries are explicit.
- Generated UI is validated and accessible.
- Retrieval quality is testable.
- Tools are permissioned, audited, and reversible when possible.
- Observability captures cost, latency, quality, and user correction signals.
Example reviewer comments
- "This tool changes customer state and must move from auto-execute to approval-required."
- "The generated component schema allows arbitrary markdown. Route it through a safe text renderer."
- "The retrieval policy needs stale-source behavior before launch."
Example ADR prompt
Write an ADR covering deterministic boundaries, registry schema, streaming protocol, RAG policy, tool risk classes, eval release gates, and production observability.
Release readiness checklist
- Registry schema and fallback tested.
- RAG fixtures pass quality threshold.
- Tool approvals and audit log verified.
- Prompt-injection cases tested.
- Cost, latency, and quality dashboards defined.