Free NCP-AAI practice test: 30 questions with answers

ReAct alternates reasoning with actions and observations, which is the standard loop for tool-using agents. Batch normalization is a training technique, pure completion has no action loop, and single retrieval does not iterate.

An orchestrator that plans and delegates to specialized workers supports decomposition and aggregation. A single prompt, a daily trigger and a cache do not coordinate multi-step work.

Memory holds conversation state and longer-term information for the agent to use. Drivers, switches and registries are infrastructure, not agent memory.

Typed tool schemas guide the model to produce valid structured arguments, and the runtime can validate them. Free text, answer-only regular expressions and manual entry do not enforce the structure during the call.

Plan-and-execute creates a plan of steps and then executes them, which improves control on complex tasks. Greedy sampling, random selection and caching do not plan.

A secrets manager with scoped identities keeps credentials out of prompts, code and training data. Prompts, public READMEs and training data expose the token.

A human approval step before destructive actions enforces the policy. Immediate execution, public tools and removed policies all violate the requirement.

Defensive parsing and explicit handling keep the agent stable and informative. Crashing, assuming success and creating duplicate tools are incorrect.

Validating results and documenting inputs and outputs keep tool use reliable. Trusting outputs without checks, sending passwords in arguments and assuming tools cannot fail are unsafe assumptions.

Compact, documented outputs with only the necessary fields are easier for the model to use reliably. Full dumps, images and removed outputs make reasoning harder or impossible.

Separation of instructions and guardrails that refuse disclosure protect the system prompt. Publishing instructions, granting admin rights and letting users edit instructions all weaken control.

Agents that read untrusted content can be redirected by hidden instructions, which is indirect prompt injection. Disk speed, GPU memory and colors are not this risk.

Scoped permissions per tool limit the damage from misuse or error. Shared admin accounts, account creation and universal write access violate least privilege.

Allow lists constrain what the agent can do, and monitoring detects and investigates unsafe behavior. Unrestricted access, high temperature and disabled review increase risk.

Topic restrictions, refusal behavior and human referral enforce the policy and direct users to appropriate help. Unlimited attempts, disclaimers only and removed checks do not meet the requirement.

Task success rate against defined criteria measures outcomes that matter to users. Token speed, tool counts and log volume measure activity or performance rather than success.

Tracing each step lets teams compare runs and locate the failing decision. Missing logs, deleted runs and guesswork make root cause analysis impossible.

A test set with expected tool choices scores both correctness and efficiency. A single unscored question, length of text and CPU temperature do not evaluate tool selection.

A regression suite compared to a baseline shows whether existing behavior holds. Deploying blindly, skipping tests and measuring only latency do not establish that behavior is preserved.

Validating tool calls against schemas in tests catches argument errors before deployment. User counts, log colors and holiday calendars do not detect argument mismatches.

The NeMo Agent Toolkit supports building and evaluating agent workflows with tools and models. Omniverse is for simulation collaboration, DeepStream for video analytics and Holoscan for medical and sensor applications.

NIM provides containerized, optimized inference endpoints with standard APIs for agents to call. Parabricks targets genomics, Earth-2 targets climate models and Riva focuses on speech.

NeMo Guardrails defines rails that constrain topics, actions and responses in conversational systems. Curator prepares data, the TensorRT builder compiles engines and cuDF processes dataframes.

NeMo Retriever provides embedding, retrieval and reranking components for RAG-based agents. Morpheus is a security pipeline framework, Isaac Sim is for robotics simulation and Jetson is hardware.

The NeMo Agent Toolkit includes profiling and evaluation for agent workflows, including step-level metrics. Omniverse, DeepStream and Parabricks address different domains.

A grounding check verifies claims against retrieved evidence, which catches unsupported citations. Higher temperature increases randomness, removing citations hides the problem, and images cannot be searched as text.

Multi-hop retrieval follows connections across documents, which answers questions that need more than one hop. Single exact searches, random samples and first-document-only retrieval miss the linked evidence.

Selecting and summarizing relevant passages keeps context focused, which reduces cost and latency. Adding everything, removing instructions and increasing output limits all make the problem worse.

Rerankers score candidates more precisely against the query, which improves the top passages the agent uses. They do not delete documents, generate content or replace embeddings.

A parameterized query tool keeps structured access safe and efficient alongside document retrieval. Raw unvalidated SQL is an injection risk, flattening databases loses structure, and blocking access prevents the required answers.