NCA-GENL Data Analysis and Visualization Practice Question
During a RAG evaluation, a data scientist computes cosine similarity between 40,000 query embeddings and 40,000 retrieved-chunk embeddings using an NVIDIA-accelerated pipeline. They then reduce the 4,096-dimensional vectors with t-SNE to 2D for a scatter plot, but the plot shows no separation between relevant and irrelevant retrievals. What is the MOST likely reason the visualization fails to reveal the retrieval quality signal?
⚠ Common exam trap
The trap here is treating t-SNE as a faithful global map of embedding space, when it only preserves local neighborhoods and can hide a label-separating direction.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
t-SNE collapses high-dimensional structure into two dimensions and its perplexity and local-neighborhood objective can obscure the few dimensions that separate relevant from irrelevant chunks.
t-SNE is a nonlinear local-neighborhood method whose output depends on perplexity and on which local structure dominates. With 4,096-dimensional embeddings, the dimensions that separate relevant from irrelevant chunks may carry little of the local variance t-SNE preserves, so the two classes interleave in the projection. A supervised or linear method that targets the label would surface the signal that t-SNE hides.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Cosine similarity cannot be computed on 4,096-dimensional embeddings, so the input matrix was invalid.
Why it's wrong here
Cosine similarity is well defined for any nonzero dimensionality, including 4,096. The computation is routine and GPU-accelerated on NVIDIA hardware. If the input were invalid the pipeline would error or produce NaNs, not a well-formed but uninformative scatter plot, so this does not explain the observed lack of separation.
- ✓
t-SNE collapses high-dimensional structure into two dimensions and its perplexity and local-neighborhood objective can obscure the few dimensions that separate relevant from irrelevant chunks.
Why this is correct
t-SNE optimizes for preserving local neighbor relationships and depends heavily on perplexity, so a signal spread across a small number of dimensions can be washed out when projecting 4,096 dimensions to two. The relevant and irrelevant points may genuinely overlap in the top local structure even though a linear probe separates them, making t-SNE the wrong tool for this retrieval-quality question.
- ✗
t-SNE preserves global distances, so the relevant and irrelevant points should be separated and the plot must be mislabeled.
Why it's wrong here
t-SNE does not preserve global distances; it preserves local neighborhoods, so global separation is not guaranteed. Claiming the plot is mislabeled ignores the algorithmic limitation. The absence of visible clusters is expected behavior when the discriminating signal lives in a few dimensions that t-SNE does not emphasize, not evidence of a labeling error.
- ✗
t-SNE is showing the raw cosine similarities rather than the embedding structure, so the relevant and irrelevant points overlap by construction.
Why it's wrong here
t-SNE operates on the embedding vectors or a distance matrix derived from them, not on the retrieval label. It never plots cosine similarity directly as a coordinate. The overlap arises because t-SNE compresses high-dimensional structure into two dimensions, not because it substitutes similarity values for embedding positions.
About these practice questions
This NCA-GENL question is part of Courseiva's 367-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official NVIDIA exam blueprint
This NCA-GENL practice question is part of Courseiva's free NVIDIA certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the NCA-GENL exam.