Skip to content

NVIDIA Certified Associate - Generative AI and LLMs (NCA-GENL) - Practice Questions

15 questions for NCA-GENL prep, weighted toward LLM fundamentals (25%), prompt engineering (20%), and RAG and vector databases (20%).

Cert page: exams/nvidia/genai-llms-associate/

Question 1

Scenario: What does the self-attention mechanism in a transformer compute?

A. The order of words in a sentence, using fixed sine and cosine position signals added to each embedding B. The model's loss, by comparing each predicted token distribution with the ground truth at the end of every step C. A weighted relationship between every token and every other token in the sequence, scored by learned query-key similarity D. The tokenizer vocabulary, by counting how often each subword appears across the training corpus and ranking them

Answer **Correct: C** **Why:** Attention lets each position attend to all others, weighting them by learned query-key similarity, which is what captures long-range dependencies. Order comes from positional encoding, not attention itself, because attention alone is permutation-invariant.

Question 2

Scenario: A model has a 128k token context window. What does that limit?

A. The combined tokens of the input plus the generated output for a single request B. The number of parameters in the model, which sets its memory footprint and is fixed once training finishes C. The training data size, meaning the number of documents the model saw before its weights were frozen D. The vocabulary size of the tokenizer

Answer **Correct: A** **Why:** The context window is the working memory for one request and covers prompt and completion together, which is why a long prompt reduces the room left for output. It is unrelated to parameter count, which is model capacity, and to the training corpus.

Question 3

Scenario: Which customization approach best teaches a model a specific output style and format with a few hundred examples?

A. RAG, which retrieves reference documents from a vector store and adds them to the prompt at inference time B. Pretraining from scratch on a new corpus of several trillion tokens, which would reset every learned weight C. Parameter-efficient fine-tuning such as LoRA D. Raising temperature, which makes sampling more random but does not teach the model any new behavior

Answer **Correct: C** **Why:** LoRA trains small low-rank adapter matrices while freezing the base weights, so it needs far less data, memory, and time than full fine-tuning, and adapters can be swapped per task. RAG injects knowledge rather than behavior, and pretraining needs orders of magnitude more data.

Question 4

Scenario: A chatbot invents a policy that does not exist in the company handbook.

A. The context window is too large, so the model fills the unused space B. The temperature is too low, which makes the model overconfident and forces it to make up missing facts C. The model is broken, so its weights are corrupted and it must be retrained from the last good checkpoint D. Hallucination: ground answers with retrieval, require citations, and measure groundedness

Answer **Correct: D** **Why:** Language models generate plausible continuations and have no built-in notion of truth, so ungrounded factual questions are the highest-risk case. Retrieval plus a requirement to cite the retrieved passage is the standard mitigation, and a groundedness metric turns it into something you can measure.

Question 5

Scenario: In a RAG pipeline, what does the embedding model do?

A. Ranks the final output by scoring each candidate answer against a set of human-written reference responses B. Converts text into vectors so that semantically similar passages sit close together in vector space and can be found by nearest-neighbor search C. Generates the answer by reading the retrieved passages and writing a natural-language reply to the user D. Stores the documents and their metadata so they can be filtered and returned when a query arrives

Answer **Correct: B** **Why:** Embeddings turn meaning into geometry, which is what makes nearest-neighbor search return relevant passages for a paraphrased query. The vector database stores and indexes those vectors, and the generative model writes the answer from what retrieval returned.

Question 6

Scenario: Which index type in a vector database trades a small amount of recall for much faster search?

A. Hash index, which maps each vector to a bucket by exact value and so only finds identical vectors B. Exact k-nearest-neighbor (flat), which compares the query to every stored vector and so is exact but not faster C. B-tree, which keeps sorted keys for range scans and gives logarithmic lookups on high-dimensional embeddings D. Approximate nearest neighbor such as HNSW or IVF

Answer **Correct: D** **Why:** ANN indexes such as HNSW build a navigable graph that finds close-enough neighbors in sublinear time. Flat search is exact but scales linearly with corpus size. B-trees and hash indexes serve exact-match and range queries, not similarity in high-dimensional space.

Question 7

Scenario: Prompt output quality improves noticeably when the model is asked to "think step by step."

A. Few-shot prompting, which supplies worked input and output examples inside the prompt B. Chain-of-thought prompting, which elicits intermediate reasoning steps C. Quantization, which lowers weight precision to save memory D. Retrieval augmentation, which fetches relevant documents and adds them to the prompt

Answer **Correct: B** **Why:** Chain of thought elicits intermediate reasoning, which helps most on multi-step arithmetic and logic. Few-shot provides examples, which is a different lever aimed mainly at format and edge cases. The two combine well.

Question 8

Scenario: Which NVIDIA offering packages an optimized model as a container with a standard inference API?

A. NVIDIA DCGM, which exports GPU health and utilization metrics from the data center for monitoring B. RAPIDS, a suite of GPU-accelerated libraries for data science C. NVIDIA NIM microservices, containers bundling an optimized model with a standard API D. Nsight Systems, a profiler that traces CPU and GPU activity

Answer **Correct: C** **Why:** NIM microservices bundle the model, the optimized runtime (typically TensorRT-LLM), and an OpenAI-compatible API into a deployable container. DCGM is GPU monitoring, RAPIDS is GPU data science, and Nsight is a profiler.

Question 9

Scenario: A model must run on a smaller GPU with acceptable quality loss.

A. Quantize the weights to a lower precision such as INT8 or FP8, trading a small accuracy loss for much less memory B. Add more layers, which deepens the network and increases its representational capacity per token C. Increase the batch size, which processes more requests per step and raises peak memory use D. Increase the context window, which lets the model read longer prompts at the cost of more GPU memory

Answer **Correct: A** **Why:** Quantization reduces the bits per weight, cutting memory footprint and often increasing throughput, at some accuracy cost that varies by method and model. Post-training quantization is the quick path; quantization-aware training recovers more quality when the drop matters.

Question 10

Scenario: What is tokenization?

A. Removing stop words and punctuation from the text so that only the meaningful terms reach the model B. Splitting text into subword units the model processes, each mapped to an integer ID C. Encrypting the prompt before it is sent so that the hosting provider cannot read the request D. Assigning API keys to each client so that the service can meter the requests that user makes

Answer **Correct: B** **Why:** Models operate on token IDs, not characters or words. Subword tokenization means an unusual word may cost several tokens, and non-English text often costs more tokens per character, which has direct pricing and context-budget consequences.

Question 11

Scenario: Which NVIDIA framework provides programmable guardrails for conversational AI applications?

A. cuDF, a GPU dataframe library that speeds up pandas-style data preparation B. Triton Inference Server, which serves models from many frameworks at scale with batching C. Omniverse, a platform for 3D simulation and digital twins built on OpenUSD D. NeMo Guardrails, a toolkit for adding programmable rails to LLM apps

Answer **Correct: D** **Why:** NeMo Guardrails defines topical, safety, and security rails in a configuration language, sitting between the user and the model. Triton serves models, cuDF is GPU dataframes, and Omniverse is the simulation and 3D platform.

Question 12

Scenario: A responsible AI review flags that the training data may under-represent a demographic group.

A. Ignore it if overall accuracy is high, since a strong aggregate score shows the model is fair to every group B. Add a disclaimer to the product page only and ship the model unchanged C. Evaluate performance per subgroup, document the limitation, and mitigate before deployment D. Increase model size, because a larger network always learns every group well

Answer **Correct: C** **Why:** Aggregate accuracy hides subgroup failure, so disaggregated evaluation is the control. Documentation (model cards) and mitigation are both expected, and a disclaimer alone shifts the burden to users without reducing the harm.

Question 13

Scenario: Which describes RLHF?

A. Reinforcement learning from human feedback, using a reward model trained on human preferences B. A vector index that organizes embeddings into a graph so that approximate nearest-neighbor queries run quickly C. A quantization scheme that stores each weight in fewer bits D. Retrieval from a large corpus of documents, which supplies grounding passages

Answer **Correct: A** **Why:** RLHF is the alignment stage that shapes helpfulness and harmlessness after pretraining and supervised fine-tuning. Direct preference optimization is a more recent alternative that skips the separate reward model.

Question 14

Scenario: Two chunking strategies are compared for a RAG system over legal contracts.

A. Always use one sentence per chunk, because the smallest unit gives the most precise embeddings for every contract clause B. Boundaries matter: too small loses context, too large dilutes relevance, and overlap avoids splitting a fact C. Always use one chunk per document, so each passage carries the whole contract D. Chunk size is irrelevant, because the embedding model normalizes any length of text

Answer **Correct: B** **Why:** Chunking is the single highest-leverage knob in most RAG systems. Structure-aware splitting on sections or clauses usually beats fixed character counts for documents with real structure, and a modest overlap protects against boundary loss.

Question 15

Scenario: How should an LLM feature be evaluated before release?

A. A held-out evaluation set with task metrics plus safety tests, run automatically on every model or prompt change B. Latency benchmarks only, since response time is the one property that cannot be checked after the release C. Manual review of a handful of prompts chosen by the developer who wrote the feature, done once on their own laptop before the release D. User feedback after launch, collected through thumbs-up buttons and support tickets once real traffic arrives at scale

Answer **Correct: A** **Why:** Evals are the regression suite for a nondeterministic system, and they only earn their keep if they run on every change: prompt edits and model upgrades both change behavior. Post-launch feedback is a lagging signal that arrives after users are affected.

Where to go deeper