Docs
View sourceArchitecture Review Exam
Timebox: 90 minutes
Scenario: You are designing an Enterprise Knowledge Copilot for architecture review. It must answer with citations from internal standards, support approved tool calls, trace every interaction, and provide a production readiness summary.
Section 1: Layering
- Draw the six major layers of the system and state the responsibility of each layer.
- Explain why the model runtime should not own retrieval policy.
- Explain why the agent/workflow layer should not own long-term experiment lineage.
Section 2: Runtime Decision
You have three candidate serving options: hosted API, vLLM, and llama.cpp.
- Create a decision matrix with at least six criteria.
- Pick one runtime for the first production release and justify the choice.
- Define the metrics that would force a runtime change.
Section 3: RAG Data Contract
- Define the required document, chunk, metadata, and query fields.
- Explain how access control should be enforced.
- Define a retrieval evaluation plan.
Section 4: LLMOps And Evaluation
- Define the trace schema.
- Define a promotion gate for prompt, retrieval, and model changes.
- Explain how MLflow-style lineage differs from LLM trace observability.
Section 5: Security And Governance
- Identify five security risks specific to this copilot.
- Define the tool governance policy.
- Define what must be logged for auditability.
Section 6: Production Readiness
- Create a release checklist.
- Define rollback behavior.
- Define three failure rehearsal scenarios.
Scoring Rubric
| Area | Points |
|---|---|
| Layering and boundaries | 15 |
| Runtime decision quality | 15 |
| RAG data contract | 15 |
| Evaluation and LLMOps | 20 |
| Security and governance | 20 |
| Production readiness | 15 |