CareerExamPro

NVIDIA Agentic AI (NCP-AAI), GPU Clusters (NCP-AII) & AI Power/Cooling practice questions

20 questions from the NVIDIA NCP-AAI, MS AI-103, GenAI Leader practice test.

Start practising

NVIDIA Agentic AI (NCP-AAI), GPU Clusters (NCP-AII) & AI Power/Cooling practice drill

10 realistic exam questions, graded as you go, with the exact rule or statute citation for every answer. Start free, no sign-up.

Sample questions

A financial research agent currently uses a single-model ReAct loop on a large frontier LLM, causing high latency and cost because it re-reasons over the entire growing conversation history after every simple API lookup. How does migrating to a Plan-and-Execute architecture address this bottleneck?

  1. A planner creates a multi-step plan upfront, smaller workers execute steps, and replanning triggers only on exceptions.
  2. It removes all tool-calling interfaces and pre-generates every analytical answer into a static cache before queries arrive.
  3. It forces the frontier model to emit all intermediate observations inside hidden reasoning tokens without calling APIs.
  4. It replaces structured reasoning with unguided zero-shot completion so every financial calculation completes in a single forward pass with fixed latency.
Show the answer

A. A planner creates a multi-step plan upfront, smaller workers execute steps, and replanning triggers only on exceptions.

Plan-and-Execute decouples high-level planning from step execution: a capable planner creates a structured step list upfront, lightweight/cheaper models execute independent steps (often concurrently), and the planner is re-invoked only when results require replanning. This sharply reduces redundant frontier-model calls and latency compared to step-by-step ReAct loops.

Reference: NVIDIA NCP-AAI Blueprint (Reasoning & Planning Loops: ReAct vs. Plan-and-Execute)

A multi-session customer operations agent needs to remember three distinct categories of state: (1) messages within the active troubleshooting thread, (2) confirmed account preferences and product entitlements, and (3) prior resolution trajectories that worked for similar outage tickets. How should these map to agent memory layers?

  1. Store active messages in semantic memory, account preferences in GPU KV-cache memory, and prior trajectories solely in the model's fine-tuned weights.
  2. Active messages in short-term working memory, account facts in semantic memory, and past trajectories in episodic memory.
  3. Concatenate all three categories inside one static system prompt shared globally across every tenant session.
  4. Store active messages in episodic memory, product entitlements in short-term scratchpad memory, and past trajectories in volatile browser cookies.
Show the answer

B. Active messages in short-term working memory, account facts in semantic memory, and past trajectories in episodic memory.

Agentic architectures separate short-term working memory (current thread state/checkpoint) from long-term memory, which subdivides into semantic memory (distilled facts, user preferences, entity graphs) and episodic memory (indexed past interaction trajectories and outcomes retrieved by task similarity to guide future decisions).

Reference: NVIDIA NCP-AAI Blueprint (Agent Memory Architecture: Working, Semantic & Episodic Memory)

A healthcare benefits assistant built with NVIDIA NeMo Guardrails must block jailbreak attempts before model inference, drop retrieved chunks containing another member's PII prior to synthesis, and mask any accidental SSN in the final response. Which combination of programmable rail stages enforces these three controls?

  1. Dialog rails for jailbreak detection, execution rails for chunk deduplication, and input rails for masking sensitive tokens in the final model output.
  2. Input rails for jailbreak screening, retrieval rails for filtering unauthorized chunks before generation, and output rails for final PII redaction.
  3. Output rails for blocking malicious prompts before inference, dialog rails for vector indexing, and retrieval rails for authenticating OAuth bearer tokens.
  4. Execution rails for prompt moderation, embedding rails for stripping PII during model pre-training, and dialog rails for network TLS termination.
Show the answer

B. Input rails for jailbreak screening, retrieval rails for filtering unauthorized chunks before generation, and output rails for final PII redaction.

NVIDIA NeMo Guardrails defines distinct interception points across the agent lifecycle: input rails inspect user prompts for jailbreaks or off-topic requests prior to LLM invocation; retrieval rails inspect and filter RAG chunks before they enter the generator context; and output rails inspect generated responses for PII leakage, policy violations, or ungrounded claims before returning text to the user.

Reference: NVIDIA NeMo Guardrails Documentation (Input, Dialog, Retrieval, Execution & Output Rails)

Other topics