NCP-AAI Practice Questions With Explanations: 20 Scenarios [2026]

A retail company is building a support assistant that must look up order status in a database, search a returns-policy knowledge base, and, when a refund is warranted, call a payments API. The latency budget is four seconds per turn, and the team wants the fewest moving parts that still handle all three capabilities reliably. Which architecture should you recommend?

You are designing a market-research system in LangGraph (a framework that models an agent workflow as a graph of nodes and edges over shared state) where a planner decomposes a brief into sub-questions, several researcher agents work concurrently, and a writer assembles the report. Reviewers report that researchers duplicate each other's work and the writer sometimes starts before all research is complete. Select TWO changes that address these problems directly.

An insurer runs six independently deployed agents (intake, fraud screening, pricing, document extraction, notification, audit). Each agent calls the others over point-to-point HTTP, and adding a compliance agent last quarter required code changes in four existing services. The architecture team wants future agents to react to relevant events without modifying existing ones. Which coordination pattern fits this requirement?

A logistics agent built with LangChain calls a reschedule_shipment tool that takes a shipment ID and an ISO-8601 date. In production, about one call in twenty fails because the model passes values like "next Tuesday" or omits the shipment ID, and the downstream API answers with HTTP 400. You want to cut these failures without changing the API. What should you do first?

A field-service agent must accept a technician's photo of an equipment nameplate together with a typed fault description, extract the model and serial number, and check warranty status through an internal API. Nameplates vary in layout, font, and lighting. Which implementation handles this most reliably inside an agentic pipeline?

Your LangGraph agent occasionally answers questions about competitor pricing (out of scope) and sometimes calls the send_email tool without the user asking. The model is a Nemotron instruct model served through NVIDIA NIM (containerized inference microservices with an OpenAI-compatible API). Before adding heavier machinery, you want fixes at the prompt and tool-definition layer. Select TWO changes that address both behaviors there.

A retrieval-augmented generation (RAG) assistant answers HR-policy questions from an internal document store. Users report confident answers that cite the wrong policy version. The team currently tracks only end-to-end answer accuracy on a 200-question golden set (a fixed, labeled evaluation dataset). Which additional metric would most directly locate the source of this failure?

You want to compare a new plan-and-execute prompt against the current ReAct prompt for a travel-booking agent. Product insists that no customer sees a degraded booking flow during the comparison, and finance wants a cost comparison measured on real production traffic. Which testing approach satisfies both constraints?

After upgrading the LLM behind a claims-triage agent, task completion on the golden set stayed flat, but the agent now selects the wrong tool on about 12 percent of multi-tool cases, up from 3 percent. Two sprints of prompt changes recovered only part of the gap. What is the most appropriate next optimization step?

A NIM-served LLM runs behind a Kubernetes deployment that autoscales on CPU utilization. During peaks, requests queue for 20 seconds while CPU stays near idle, so no new replicas start. GPU nodes need about 90 seconds to become ready. Which change most directly fixes the missed scale-out?

You are shipping a new version of a stateful customer-service agent whose conversation state is checkpointed to Postgres through a LangGraph checkpointer (a component that persists graph state after each step so a thread can resume). The new version changes the state schema. Operations wants zero dropped conversations and a fast rollback path if quality regresses. Select TWO practices that meet both requirements.

A RAG agent stack runs three models on one eight-GPU node: a 70B-class LLM, an embedding model, and a reranker. The LLM saturates its GPUs at peak while the embedding and reranking services sit mostly idle on their own dedicated GPUs. You must raise LLM throughput without buying hardware. Which change is the best first step?

An agent generates weekly staffing schedules subject to labor rules, employee preferences, and coverage minimums. Early attempts commit to a first assignment and later hit dead ends, producing invalid schedules. Latency is flexible (minutes are acceptable) and correctness matters more than token cost. Which reasoning pattern best fits this problem?

A concierge agent holds multi-week conversations with returning travelers. Two problems appear: long threads exceed the model's context window mid-trip, and preferences a traveler stated in March (aisle seat, vegetarian meals) are forgotten by June. The team wants both fixed with minimal added latency per turn. Which memory design should you adopt?

Engineers query a RAG assistant over 40,000 pages of equipment manuals. Dense semantic retrieval performs well on descriptive questions but misses exact part numbers such as "XR-2210-B", returning chunks about similar-looking parts instead. Top-k is 5 and there is no reranking stage. Select TWO changes that most directly improve retrieval for these exact-match queries.