Practice Exam NCA-GENL: Certified Associate: Generative AI with LLMs

Question bank: 100 · timed mock exam: 60 min

Official exam page on the vendor's site (opens in a new tab)

Exam guide: format, domains and how to prepare

Compare: NCA-GENL vs Databricks Generative AI Engineer Associate

  • Core Machine Learning and AI Knowledge 30 questions
  • Software Development 24 questions
  • Experimentation 22 questions
  • Data Analysis and Visualization 14 questions
  • Trustworthy AI 10 questions
Study mode (free) Exam mode (subscribers)

Free: 25 questions in study mode, all domains, no account.

Subscription: the 100 questions of this certification, the timed exam mode, and all other certifications too.

Subscribe · €9.99 / month

Secure card payment (Stripe) · No commitment · Cancel online

How to prepare for Certified Associate: Generative AI with LLMs with InfraPrep

The 100 questions cover 5 domains: Core Machine Learning and AI Knowledge, Software Development, Experimentation, Data Analysis and Visualization, Trustworthy AI. 25 of them are free, without an account.

Study mode

Answer at your pace: each answer is corrected straight away, with an explanation and the link to the documentation it relies on.

Exam mode

A timed exam of 60 minutes, with a question map to flag and come back to questions before you submit.

Results by domain

Your score is broken down by domain, so you know what to review first.

Original questions

Every question is written from the public documentation, never copied from the exam. You can contest an answer from study mode.

10 free sample questions

Question 1 Core Machine Learning and AI Knowledge

A telecom operator wants a support chatbot that answers questions from its product manuals, which are updated almost every day. The team has no budget for repeated training runs and wants answers to reflect the latest manuals. Which approach best fits these requirements?

  1. Retrieval-augmented generation over the current product manuals
  2. Full fine-tuning of the model every night on the updated manuals
  3. A larger pretrained model with a longer original training period
  4. A higher sampling temperature so the model generalizes more freely
Show the answer and the explanation

Correct answer: A

NVIDIA describes RAG as adaptable to a new knowledge base without retraining or fine-tuning the generative model, and as keeping answers current by pulling in new data as it appears.

Source: official documentation

Question 2 Data Analysis and Visualization

An analyst explores several gigabytes of chatbot logs with an existing pandas notebook (groupby aggregations, joins, filters). The notebook has become slow, and the team does not want to rewrite the analysis code. What should the analyst do on a GPU workstation?

  1. Enable the cuDF pandas accelerator and keep the pandas code
  2. Rewrite the notebook with nested Python loops over NumPy arrays
  3. Randomly drop most of the log rows before running the analysis
  4. Export the logs to CSV files and analyze them in a spreadsheet
Show the answer and the explanation

Correct answer: A

cuDF accelerates pandas with zero code change: operations run on the GPU when possible and fall back to pandas on the CPU otherwise.

Source: official documentation

Question 3 Experimentation

In a RAG evaluation, retrieval precision and recall are good and answers stay on topic, but reviewers find answers that state facts contradicting the retrieved passages. Which metric should the team track to measure this specific problem?

  1. Response relevancy
  2. Retrieval recall
  3. Faithfulness
  4. BLEU score
Show the answer and the explanation

Correct answer: C

Faithfulness measures the factual consistency of the generated answer against the given context, which is exactly what contradicting the passages violates.

Source: official documentation

Question 4 Software Development

A stateless embedding model served by Triton Inference Server receives thousands of small, independent requests per second. GPU utilization stays low and throughput is below target. Which configuration change should be tried first?

  1. Enable the dynamic batcher in the model configuration
  2. Switch the model to the sequence batcher for each client
  3. List many preferred batch sizes for the model by default
  4. Set the maximum batch size to zero to disable batching
Show the answer and the explanation

Correct answer: A

Dynamic batching lets Triton combine inference requests into batches on the server, which typically increases throughput, and it is the batcher intended for stateless models.

Source: official documentation

Question 5 Trustworthy AI

Publishing a model card that describes the training datasets, training methods, performance measures, and ethical considerations of a model mainly supports which principle of trustworthy AI?

  1. Transparency
  2. Nondiscrimination
  3. Privacy
  4. Safety and security
Show the answer and the explanation

Correct answer: A

NVIDIA places model cards under transparency, the principle that helps stakeholders understand how an AI model was trained and how it works.

Source: official documentation

Question 6 Core Machine Learning and AI Knowledge

A RAG assistant returned relevant passages for months. Last week an engineer switched the query-side embedding model to a newer model to improve accuracy, but left the existing vector index, built with the old model, untouched. Retrieval now returns mostly unrelated chunks. What is the most likely root cause?

  1. The LLM context window is now too small for the retrieved chunks
  2. Query vectors and stored vectors now come from different embedding models
  3. The vector database needs more memory to serve the larger new vectors
  4. The chunk overlap used at ingestion time is now too low for the new model
Show the answer and the explanation

Correct answer: B

Retrieval compares the query embedding with the stored document embeddings, so the documentation stresses keeping ingestion-time and query-time embeddings consistent; mixing two models breaks that comparison.

Source: official documentation

Question 7 Data Analysis and Visualization

A team compares two embedding models for a RAG pipeline. The pipeline passes the top five retrieved chunks to the LLM, which reads all five regardless of their order. Which retrieval metric best reflects what matters for this pipeline?

  1. nDCG@10, because it rewards the ideal ranking order
  2. Recall@5, because only presence in the top five matters
  3. BLEU, because it compares outputs with reference texts
  4. Perplexity, because it measures the model's uncertainty
Show the answer and the explanation

Correct answer: B

Recall is rank-agnostic: it measures the share of relevant results retrieved regardless of order, and NVIDIA recommends recall@5 for most enterprise RAG pipelines.

Source: official documentation

Question 8 Experimentation

According to NVIDIA's guidance on LLM evaluation, which approach is particularly suited to judging qualities such as coherence and creativity, where automated metrics fall short?

  1. LLM-as-a-judge
  2. ROUGE overlap
  3. Perplexity
  4. Precision@k
Show the answer and the explanation

Correct answer: A

The LLM-as-a-judge approach uses an evaluator LLM with explicit criteria and is described as excelling where automated metrics fall short, such as coherence and creativity.

Source: official documentation

Question 9 Software Development

A team serves a BF16 Llama model and wants to cut its GPU memory needs and raise inference throughput. No new training run is allowed; only a small calibration dataset is available. Which NeMo capability fits?

  1. Supervised fine-tuning on a larger labeled dataset
  2. Post-training quantization to a format such as FP8
  3. A wider beam search at generation time
  4. Activation recomputation during model training
Show the answer and the explanation

Correct answer: B

NeMo post-training quantization converts a BF16 or FP16 model to FP8, INT8, or INT4 for efficient serving, with a lightweight calibration step on a small dataset; its benefits include lower memory needs and higher throughput.

Source: official documentation

Question 10 Trustworthy AI

A bank has deployed an LLM assistant for account questions. Users increasingly ask it about politics and cooking, and it answers. The bank wants the assistant to stay on banking subjects. Which type of NeMo Guardrails rail addresses this?

  1. Security guardrails for third-party API calls
  2. Safety guardrails on language and data sources
  3. Topical guardrails that keep to specific subjects
  4. Federated learning across the bank's branches
Show the answer and the explanation

Correct answer: C

Topical guardrails ensure that chatbots stick to specific subjects, which is the requirement here.

Source: official documentation

Frequently asked questions

Are there free questions, and do I need an account?

Yes. 25 questions of this certification are free in study mode, across all its domains, without a subscription and without registration. The account is created when you subscribe, at checkout.

How does the subscription work, and can I cancel?

One monthly subscription covers all certifications and the timed practice exams. There is no minimum period. You cancel online, at any time, with the "Cancel your contract" function in My account; cancellation takes effect at the end of the month already paid, and no further payment is taken.

Where do the questions come from?

Every question is original, written from the vendor's public documentation. Each answer comes with an explanation and a link to the documentation page it relies on. InfraPrep is not affiliated with the vendors, and certification names belong to their owners.

In which language are the questions?

The questions are written in English, also on the French version of the site.

Is the catalog updated?

Yes, the catalog changes over time: new certifications and new questions are added.