Databricks Certified Context Engineer Associate Practice Tests

Explanation

MCP and RAG are complementary, not competing. RAG excels for stable, evergreen content (metric definitions) via indexed retrieval with pre-computed embeddings. MCP is designed for live, transactional data (stock prices refreshed every 5 minutes) where freshness is critical and pre-indexing is impractical. Using RAG for both creates staleness problems. Using MCP for definitions wastes tokens on stable content. The principle is match the tool to the data freshness requirement.

Explanation

MLflow 3 provides tracing and evaluation frameworks with built-in scorers (answer_correctness, context_recall, faithfulness) that quantify agent performance across configurations. MLflow traces show exactly which context the model received during inference, enabling precise measurement of configuration trade-offs and token-efficiency gains. Lakebase stores agent state but does not measure answer quality. Unity Catalog metric views define KPI calculations but are not evaluation frameworks. MCP introspection measures tool schema sizes, relevant to context analysis but not answer correctness or faithfulness. The scenario explicitly requires quantitative comparison of two configurations—MLflow 3 is the evaluation platform.

Explanation

The workflow’s required state mechanism is checkpointed agent-session/workflow state: it must preserve active execution across an interrupt for human review and enable recovery after an application restart. The supplied official Databricks corpus states that LangGraph interrupt() pauses the workflow and that checkpoint-based pause/resume enables recovery when the application restarts while awaiting approval. A Delta-backed operational object may hold purchase-order or approval data and be queried by external systems, but it is not, by itself, the workflow checkpoint/recovery strategy. The question conflates operational-data queryability with execution-state durability; among the choices, C directly addresses the stated interruption and restart requirements.

Explanation

Reasoning mode should match task complexity and token budget justification. Type A (simple lookups) requires minimal reasoning overhead; reduced thinking is economical and sufficient. Type B (multi-step financial forecasting) genuinely requires deeper analysis; standard or extended thinking is justified by task complexity. Applying extended thinking universally wastes tokens on unnecessary reasoning for Type A (75% of traffic). Applying reduced thinking to Type B risks incomplete analysis and incorrect forecasting. Standard thinking for all is a compromise but misses the selective optimization opportunity. The scenario explicitly provides reasoning budget differences (200 vs. 500 tokens) and asks for optimization—matching mode to task directly addresses both efficiency and quality constraints.

Explanation

Progressive disclosure's core principle is staged information entry. Discovery phase (tool names/summaries) is always loaded—this provides the agent with awareness of available options at minimal cost. Planning/execution phases (full schemas) load just-in-time when the agent determines which tools are relevant. "Precompute which 5 tools are most likely needed and load only their full schemas upfront; omit the others entirely" abandons discovery coverage—if an unexpected tool is needed, the agent cannot access it. "Generate a summary prompt describing all 18 tools in natural language without formal schemas to fit more context" loses the semantic detail required for accurate tool selection. "Distribute tool schemas across the conversation to spread token load evenly" (distributing schemas) does not reduce baseline cost; it merely defers the token consumption. "Load the names and one-line summaries of all 18 tools in context at session start; retrieve full schemas only for tools the agent identifies as relevant during task execution" implements the staged approach correctly.