NCA-AMAD Practice Questions: 20 Exam-Style Questions Explained (2026)

ReAct runs a loop of thought, action and observation, so the agent sees each result, including a timeout, before it decides what to do next. ReWOO plans every step up front, which saves tokens on stable chains, and the NeMo Agent Toolkit documentation notes that it cannot adapt to unexpected tool failures or new information. A router agent picks one branch in a single pass and does not chain dependent lookups. Pasting every supplier record into one call skips the lookups entirely and makes the prompt grow with each supplier added.

NeMo Guardrails execution rails validate tool calls, their arguments and their results, so a refund that breaks policy can be blocked whatever the email said. A human approval step puts a person in front of an irreversible action, and the NeMo Agent Toolkit can pause a workflow until a user responds. A larger model may follow instructions more closely, but it still reads the attacker's text as input and gives no guarantee. Longer log retention helps an investigation after the fact and stops nothing. Temperature changes how varied the wording is and has no effect on whether injected instructions are followed.

NVIDIA's glossary separates short-term memory, the context of the current workflow, from long-term memory, knowledge that persists such as a user's preferences. The NeMo Agent Toolkit memory module stores conversation history, preferences and long-term memory with providers such as Mem0, Redis or Zep. Replaying every transcript inflates the prompt and the cost of every turn. Nightly fine-tuning is slow and expensive for per-customer facts and mixes customers' data into shared weights. A larger context window still forgets everything when the session ends.

NVIDIA's agent evaluation guidance tracks trajectory efficiency, the steps or tokens spent per successful task, and recommends explicit budgets such as a share of tasks finished under a set number of tokens and tool calls. Watching it next to the success rate shows whether the agent got cheaper without getting worse. The success rate on its own hides the cost. Answer length says little about the steps behind the answer. GPU utilization describes the servers, and it moves with traffic from other users.

Semantic deduplication embeds the documents, clusters them and compares pairs within a cluster by cosine similarity, which catches paraphrases and translations that share meaning but few characters. NVIDIA's text processing guidance describes exact deduplication as fast and blind to near-variants, and fuzzy deduplication with MinHash and Jaccard similarity as the method for minor edits, partial overlaps and formatting differences. Language identification would drop every German and Japanese article, including those with no English counterpart.

NVIDIA describes a data flywheel as a loop in which data from AI interactions, including incorrect predictions and low-confidence outputs, is processed, used to customize the model and evaluated before the improved model goes back into production. The new product names are missing from the original training set, so retraining on it repeats the problem. A larger model has not seen the names either. Stopping the logs throws away exactly the data that shows what changed.

NVIDIA's text processing pipeline includes redacting personal information (masking it, substituting non-sensitive equivalents, or removing it) and deduplication, and its data flywheel description removes personal information and noise before customization. Keeping logs raw would carry personal data into the model's weights. Copying common questions creates duplicates that skew training toward them. Nothing in NVIDIA's pipeline requires translation, and language identification usually comes early so that later filters can run per language.

NVIDIA's benchmarking guidance defines time to first token as queuing time plus prefill time plus network latency, and notes that when concurrency exceeds the maximum batch size multiplied by the number of active replicas, requests wait in a queue and time to first token rises. The steady time between later tokens rules out slower decoding. Nothing in the scenario points to truncated prompts, and a slow network would show at low traffic too. More replicas or more capacity per replica shorten the queue.

TensorRT-LLM can keep the KV cache blocks of completed requests and reuse them for a new request that starts with the same prefix, so the 6,000 shared tokens skip prefill on later calls. Speculative decoding speeds up the generation of output tokens and leaves prefill as it was. A larger batch raises throughput, and every request in it still runs prefill on the full prefix. Streaming shows tokens as they are generated, while the GPU still processes the full prefix for every request.

TensorRT-LLM documents speculative decoding as a technique for accelerating inference at low batch sizes: a lightweight drafting mechanism proposes candidate tokens, the target model verifies them in a single forward pass, and the tokens that match are accepted, which cuts the number of sequential passes. A larger batch raises throughput when many requests are waiting, which this service does not have. Disaggregated serving helps when long prefills interfere with ongoing decodes at scale. Quantizing to 4 bits would change the model's precision, which the team wants to keep.

NVIDIA's NIM guidance describes two ways to deploy LoRA: merge an adapter into the base model, which adds no latency but serves one task per deployment, or load adapters dynamically on one base model, where each request names its adapter and a batch can mix adapters. With many adapters and light traffic, dynamic loading shares one deployment. Forty merged copies would tie up forty GPUs that sit mostly idle. Folding every adapter into one set of weights blends the customers' behavior together, and a full fine-tune on pooled data does the same while mixing their data.

NVIDIA's RAG Blueprint documentation describes hybrid search as combining sparse keyword matching (BM25) with dense semantic embeddings, and recommends it for documents with domain-specific technical terms. Keyword matching is what catches an exact code. Larger chunks surround the part number with more text and leave the retrieval method unchanged. A bigger embedding model still encodes meaning, and part numbers that differ by one character mean almost the same thing to it. Temperature affects generation after retrieval has already returned the wrong passages.

NVIDIA's RAG guidance recommends starting with RAG for relevant, up-to-date information and notes that a RAG system can cite its sources. A weekly amendment becomes searchable as soon as it is indexed. Fine-tuning, full or LoRA, puts knowledge into weights, needs a new training run for each change, and gives no pointer to a source clause. Few-shot examples shape the format of answers but give the model no access to the contracts.

NVIDIA's NeMo Retriever reranking NIM scores a query against each candidate passage and reorders the list by how well they match, so relevant passages can move into the top positions the LLM sees. More embedding dimensions change how the first retrieval encodes text, and the reranker is the step designed to order candidates the retriever has already found. Sending all 50 passages raises the prompt size and the cost of every question. A guard model screens content for safety and does not rank relevance.

NVIDIA's RAG Blueprint documentation says reasoning improves accuracy on complex multi-step questions, logical deduction and technical problems, at the cost of extra latency from reasoning tokens, and lets you turn thinking on or off. Turning it off everywhere gives up the gains on the hard questions. A larger reasoning budget lets the model think longer, which adds time. A larger model usually generates each token more slowly.

NVIDIA AI Enterprise's Long-Term Support Branch is meant for highly regulated industries: a new one arrives about every two and a half years, each is supported for three years, and fixes ship quarterly. The Feature Branch brings the newest features with about a month of support. The Production Branch offers API stability for mission-critical work, but nine months of support would force upgrades during the period. The Infrastructure Branch covers the drivers and operators underneath the AI software.

NVIDIA's VSS blueprint ingests live or archived video and provides summarization and interactive question answering with a vision-language model, which is the job described. Transcribing the video keeps what is said and loses what the cameras show. The data flywheel blueprint improves models from production traffic and does not search video. The digital human blueprint is a conversational front end.

NVIDIA Cosmos world foundation models generate and transform physically grounded video, and NVIDIA lists synthetic video data and more varied weather, lighting and locations among their uses. Nemotron models work with language and can describe a scene without producing new video. Isaac GR00T is a foundation model for humanoid robot skills. Riva handles speech recognition and speech synthesis.

NVIDIA's guide to autoscaling NIM on Kubernetes uses gpu_cache_usage_perc, the KV cache utilization NIM exposes on its Prometheus endpoint, fed to the Horizontal Pod Autoscaler through Prometheus and the Prometheus Adapter. It rises with the number of requests the GPU is handling, which is where LLM load lives. CPU utilization stays low while the GPU is busy, so a lower CPU target would still track the wrong resource. The allocated GPU count is fixed by the specification, and pod memory on the node does not follow how busy the GPU is.

MIG, available on Ampere and later GPUs, splits one GPU into up to seven instances with their own memory, compute and fault isolation, which gives each customer predictable performance. Time-slicing shares the GPU in turns with no memory or fault isolation between the replicas. MPS runs kernels concurrently, and a fatal fault in one client can bring down the others. Kubernetes allocates whole GPUs to pods by default, so three pods that each request a full GPU would need three GPUs.