Independent: not affiliated with NVIDIA. Free to try; the study bundle adds a study guide and two practice exams.
5 original questions, none of them in the practice exams.
Pick an answer. Show answer explains every option and links its NVIDIA source.
5 questions / Explanations link NVIDIA sources / Last updated Oct 26, 2026
Question 1
Multiple choice
Domain 1 / Task 1.9 Apply prompt engineering
An extraction service asks an LLM to return each invoice's supplier, date and total as JSON for a downstream program. About one response in fifty is not valid JSON, and the program fails on it. Which change addresses this most directly?
Show answerHide answer
Answer: B
Why
Guided (constrained) decoding enforces a schema while the model generates, which an instruction or examples alone cannot guarantee. In a NIM, the bundled vLLM or SGLang engine applies it when the request carries the schema.
Why not the others
ATemperature makes sampling more random. It does nothing to hold a format.
CExamples help with format but do not enforce it, and they cost tokens on every call.
DEncoder-only models suit labels, spans and embeddings, not generated JSON. Context length is not the problem.
Domain 2 / Task 2.6 Evaluate and refine models and benchmarks
A team is tuning a RAG assistant over internal wiki pages. There are no reference answers, but each logged response is stored with the passages that were retrieved for it. The team wants to measure how often answers make claims the passages do not support. Which measure fits?
Show answerHide answer
Answer: C
Why
Faithfulness (groundedness) asks whether an answer is supported by the retrieved context. It needs that context, which the team logs, and no reference answer, which it lacks.
Why not the others
ACounts word overlap, so a faithful paraphrase scores low. Overlap does not show support.
BAn intrinsic language-model measure. It says nothing about whether a claim is supported.
DA general knowledge benchmark. It never looks at this assistant's answers.
Domain 1 / Task 1.8 Choose and use text embedding models
A team is choosing an embedding model for a RAG index of Spanish-language insurance policies, split into chunks of about 1,000 tokens. Which two checks belong in the choice? Choose two.
Show answerHide answer
Answers: A and D
Why
Two of the four checks for an embedding model: it covers the language and domain of the content, and its maximum input length fits the chunks. The other two are a dimension that fits the storage and latency budget, and good ranking on a retrieval benchmark for similar data.
Why not the others
BThe embedding model and the LLM need not share a vocabulary. What must match is the embedding model used for queries and for documents.
CDimension should fit the storage and latency budget. Larger is not a requirement.
EPreference alignment shapes how a generator responds. It is not one of the checks for an embedding model.
Domain 3 / Task 3.4 Identify the data, hardware and software components needed
A platform team must serve one very large LLM across several multi-GPU nodes. Prefill and decode should run on separate workers, and each request should go to a worker that already holds its KV cache. Which NVIDIA component is built for this?
Show answerHide answer
Answer: C
Why
Dynamo is NVIDIA's distributed inference framework over TensorRT-LLM, vLLM and SGLang. It splits prefill and decode across workers, routes requests with KV-cache awareness, and autoscales with a planner.
Why not the others
AThe renamed Triton Inference Server: a model server for many models and frameworks, with dynamic batching. Dynamo is a separate NVIDIA framework for LLM serving across many nodes.
BRuns fine-tuning jobs through a REST API. It does not serve models.
DShrinks a trained model before deployment with quantization, pruning and distillation.
An analyst has 768-dimensional embeddings for 40,000 support tickets and wants one picture that shows whether tickets from the same product line sit together. Which chart fits?
Show answerHide answer
Answer: B
Why
To see how high-dimensional embeddings cluster, project them to two dimensions with PCA, t-SNE or UMAP and plot the points. Coloring by product line shows whether tickets with the same label group together. RAPIDS cuML runs PCA and UMAP on the GPU.
Why not the others
ACompares counts across product lines. It does not show how the tickets relate.
CShows the spread of one variable, not whether tickets group.
DForty thousand rows of raw values. Raw dimensions do not show which tickets sit near each other.
The study bundle has two practice exams with 120 original questions weighted to the five exam domains, an answer key and an explanation for every answer, plus the 60-page study guide. Practice exam explanations are shorter than the ones on this page: a short paragraph per question, with no source links.
Study bundle
$19
120 practice questions / Study guide / Answer keys
Every bundle includes 6 months of access: download the PDFs, and any new editions, until 6 months after your purchase date. Your receipt has a license key. Enter it at overqual.com/download for the three PDFs, stamped with your email and order number. They are yours to keep.