Free NCP-AAI: NVIDIA Certified Professional — Agentic AI Exam Questions and Answers

Which two statements about embedding models in a RAG pipeline are correct? (Choose two.)

Select 2 answers.

• AA. Replacing the embedding model of an existing index requires re-embedding only the incoming user queries
• BB. Query vectors must come from the same embedding model that produced the indexed document vectors
• CC. Many retrieval embedding models accept an input type setting so queries and passages are encoded differently
• DD. Vectors from two different embedding models can share one index when both are normalized to unit length

An LLM NIM runs as a Kubernetes Deployment. Each time a pod restarts, it takes about 12 minutes to pass its readiness check, and logs show the container downloading model files from NGC before it starts serving. During a node failure, the service is unavailable for that whole period. What should the engineer do?

• AA. Increase the readiness probe initial delay so Kubernetes stops restarting the pod early
• BB. Mount a persistent volume at the NIM cache path so downloaded model assets are reused after a restart
• CC. Lower the maximum batch size so that the engine initializes with less GPU memory
• DD. Raise the pod's CPU request so the model files download and load faster

A team split a contract-review assistant into five agents (extractor, classifier, risk scorer, summarizer, formatter) that run strictly in sequence. Each agent needs the full output of the previous one, and a prompt change in one stage often breaks the next. Compared with the earlier single-agent version, latency and token cost have roughly tripled with no accuracy gain. What should the team do?

• AA. Add a sixth supervisor agent to coordinate the handoffs between the existing agents
• BB. Give each agent its own vector store so that stages stop sharing context
• CC. Run all five agents in parallel and combine their outputs at the end
• DD. Merge the stages into one agent that calls the extraction and scoring steps as tools

A parts-lookup agent uses a dense-embedding retriever over 60,000 catalog chunks. Queries that describe symptoms work well, but queries containing an exact part number such as "KX-4410-B" often return chunks for similar-looking numbers instead. What should the engineer do?

• AA. Re-chunk the catalog at 2,000 tokens per chunk so each part number sits in a larger context
• BB. Add BM25 keyword search next to the dense retriever and merge both rankings with reciprocal rank fusion
• CC. Apply a cross-encoder reranker to the same four dense candidates the retriever already returns
• DD. Increase top-k on the dense retriever from 4 to 40 and pass every returned chunk to the LLM

Which retrieval metric is the proportion of all relevant documents for a query that appear in the top k results?

• AA. Precision@k
• BB. Recall@k
• CC. Normalized discounted cumulative gain (NDCG)
• DD. Mean reciprocal rank (MRR)

Every request to an agent begins with the same 3,000-token system prompt and tool definitions, followed by a short user message. Profiling shows that prefill time dominates time to first token (TTFT). Which two changes reduce TTFT? (Choose two.)

Select 2 answers.

• AA. Place the current timestamp and user ID at the start of the system prompt
• BB. Disable in-flight batching so each request prefill runs alone
• CC. Keep the system prompt and tool definitions identical and first, with per-request content after them
• DD. Enable KV cache reuse (prefix caching) in the serving engine

A developer is registering an order-lookup function with an LLM so that an agent can call it through function calling. Which two items are normally supplied to the model as part of the tool definition? (Choose two.)

Select 2 answers.

• AA. The source code of the function that implements the lookup
• BB. A description of what the tool does and when it should be used
• CC. A JSON schema that describes the tool's input parameters
• DD. The API key that the function uses to reach the order database

A data team wants an agent to investigate why weekly revenue dropped. Which query to run next depends on what the earlier queries returned, and nobody can predict how many steps an investigation will take. The team will cap each run at 15 tool calls. Which agent architecture should the team use?

• AA. A router that sends the question to one of several specialist agents with fixed query sets
• BB. A plan-and-execute agent that writes all queries up front and then runs them without revision
• CC. A prompt chain that runs five predefined SQL queries in order and summarizes the output
• DD. A ReAct-style loop in which the model chooses each next query after reading the latest result

A team embeds documents and user questions with an NVIDIA NIM embedding microservice that serves an asymmetric retrieval model. Both ingestion and search requests currently use identical settings. Retrieval relevance is poor even though the documents are chunked sensibly. Which change should the team make?

• AA. Set truncate to NONE for documents and request a smaller output dimension for questions
• BB. Set input_type to passage for documents and to query for user questions
• CC. Normalize the document vectors and leave the question vectors unnormalized
• DD. Set input_type to query for both document ingestion and user search

A trip-planning agent writes a complete eight-step plan before acting and then executes the steps in order. In one run, the tool result for step 3 shows that the chosen hotel is fully booked. The agent still runs steps 4 to 8, which depend on that hotel, and returns an itinerary with an unavailable booking. Which change BEST addresses the root cause?

• AA. Extend the planning prompt to list every possible failure case
• BB. Compare each tool result with the plan and replan when they conflict
• CC. Raise the planner temperature to produce more varied plans
• DD. Merge all eight steps into a single itinerary-generation prompt