AWS AI Practitioner Fundamentals of GenAI: 336 practice questions

12 of the 336 Fundamentals of GenAI questions in the Certsqill AWS AI Practitioner bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for AWS AI Practitioner? Take the free 5-min readiness check →

1. Tokens are model-dependent units and do not correspond: Select TWO statements that correctly guide this assess

Select two. More than one option is correct — every correct one is ticked below.

1. Tokens are model-dependent units and do not correspond universally to complete words. ✓

   Tokenization varies by model, language, and text, so counting words alone cannot reliably estimate token usage.
2. Every token represents exactly one character across foundation models.

   Token boundaries vary, and a token may represent part of a word, a word, or multiple characters.
3. Input and output tokens can both contribute to model usage charges. ✓

   Usage pricing may account for tokens sent to and generated by a supported foundation model.
4. A token always represents one complete English word for business reporting.

   Common words may be single tokens, but other words, symbols, and languages can use multiple tokens.
5. Output tokens are excluded when estimating generative model consumption.

   Generated output contributes to usage measurement when the selected model and pricing model count output tokens.

Tokens are model-dependent text units, and both input and output token usage can affect generative AI costs.

2. Represent documents and queries as numerical vectors: Select TWO requirements that embeddings can support in t

Select two. More than one option is correct — every correct one is ticked below.

1. Represent documents and queries as numerical vectors for comparison. ✓

   Embeddings encode content numerically, allowing systems to compare representations of documents and queries.
2. Rank content by semantic similarity despite different wording. ✓

   Similarity comparisons can identify conceptually related content even when query and document vocabulary differs.
3. Encrypt document contents automatically through vector conversion.

   Embedding vectors support comparison but do not inherently provide encryption or replace data-protection controls.
4. Guarantee that the highest-similarity document contains verified facts.

   Similarity reflects representation proximity, not factual accuracy, source authority, or verification status.
5. Grant users permission to view every similar document.

   Embeddings do not establish authorization; access decisions require separate identity and permission controls.

Embeddings numerically represent content and enable semantic similarity comparisons, but they do not verify facts, encrypt data, or grant access.

3. Use retrieval units that balance focus and context: Which approach is best?

1. Split every sentence into a separate retrieval unit regardless of meaning.

   Overly small units can separate related ideas and remove context needed to interpret answers.
2. Use retrieval units that balance focus and context. ✓

   Effective chunking creates focused retrieval units while preserving enough surrounding information to answer questions accurately.
3. Place every manual into one retrieval unit to preserve all context.

   Very large units can reduce retrieval focus and consume available context with irrelevant material.
4. Remove surrounding context from each passage to minimize token usage.

   Removing context may reduce usefulness when meaning depends on definitions, conditions, or nearby explanations.

Chunking should balance focused retrieval with sufficient context, because both oversized and undersized units can reduce usefulness.

4. Attention weighs relationships among sequence elements: Which concept best explains this capability?

1. Diffusion iteratively removes noise from an initially noisy representation.

   Denoising describes diffusion generation, not how transformers relate sequence elements during processing.
2. Chunking divides long documents into smaller retrieval units.

   Chunking organizes source material for retrieval but does not explain token-to-token relationships inside a model.
3. Embedding conversion grants a model access to every source document.

   Embeddings create numerical representations for comparison and do not inherently provide unrestricted document access.
4. Attention weighs relationships among sequence elements when processing the input. ✓

   Attention helps transformers consider relationships between tokens, including information located elsewhere in the sequence.

Transformer attention helps a model weigh relationships among tokens, allowing relevant information from different sequence positions to interact.

5. Iterative denoising progressively transforms noise into: Which concept best fits?

1. Iterative denoising progressively transforms noise into a generated result. ✓

   Diffusion models generate content through repeated denoising steps that move toward a structured output.
2. Chunking separates source documents into retrieval-sized passages.

   Chunking prepares information for retrieval and does not describe how diffusion generates images.
3. Attention assigns importance among tokens in an input sequence.

   Attention models relationships among sequence elements but does not define diffusion image generation.
4. Embeddings convert content into vectors for similarity comparison.

   Embeddings support numerical similarity tasks but do not describe iterative image creation from noise.

Diffusion models generate outputs by repeatedly denoising an initially noisy representation until a structured result emerges.

6. Multimodal processing for image input and text output: Which capability is required?

1. Diffusion generation for producing images from noisy representations

   Diffusion may generate images, but it does not by itself satisfy image inspection with written explanations.
2. Text-only processing for written inspection instructions

   Text-only processing cannot directly analyze image inputs required for visual inspection.
3. Semantic embeddings for comparing written maintenance terms

   Embeddings can compare representations but do not necessarily perform visual inspection or generate explanations.
4. Multimodal processing for image input and text output ✓

   A multimodal model can process more than one modality and may generate a different supported modality.

Multimodal capability is needed when an application must handle image input and produce supported text output.

7. Reuse and evaluate a pretrained foundation model: Which approach best fits this requirement?

1. Use embeddings as the organization’s language generator

   Embeddings support representation and similarity tasks but do not themselves provide general language generation.
2. Use deterministic rules for every language task

   Rules suit explicit logic but are generally unsuitable for broad, varied language generation.
3. Train a model from randomly initialized parameters

   Training from scratch requires substantially more data and effort than reusing pretrained capabilities.
4. Reuse and evaluate a pretrained foundation model ✓

   A pretrained foundation model offers broad reusable capabilities that can be assessed for specific experiments.

Reusing a pretrained foundation model provides broad capabilities without building a model from scratch.

8. Select, summarize, or chunk information when the full: Select TWO requirements that address context-window con

Select two. More than one option is correct — every correct one is ticked below.

1. Select, summarize, or chunk information when the full history cannot fit. ✓

   Managing supplied context preserves the most relevant information when documents or conversations exceed available capacity.
2. Replace access controls with a larger context window for sensitive policies.

   Context capacity does not authorize access or protect sensitive information from inappropriate inclusion.
3. Keep the combined usable input and output within the selected model’s context limit. ✓

   The context window limits the usable amount of input and output for a particular model interaction.
4. Assume a larger context window guarantees more accurate answers.

   Additional capacity may help, but larger context does not guarantee improved accuracy or relevance.
5. Treat all conversation history as equally important for every request.

   Older or irrelevant history can consume context without improving the current response.

Context windows limit usable interaction content, so teams must fit and prioritize relevant information rather than assume more context ensures accuracy.

9. Compare input and output token usage: Which cost consideration is most important?

1. Assume prompt length is free after model selection

   Input processing can contribute to usage costs depending on the model and pricing arrangement.
2. Compare input and output token usage ✓

   Both long prompts and generated responses can contribute to model usage, so output-only comparison is incomplete.
3. Assume shorter responses mean lower total cost

   Short outputs can accompany lengthy inputs, and pricing depends on the selected model and arrangement.
4. Compare only generated word counts

   Words do not reliably represent model tokens and omit potentially chargeable input context.

Compare model-dependent input and output tokens because both prompt and response length can affect usage costs.

10. Select and organize relevant context for each request: Which practice best fits?

1. Select and organize relevant context for each request ✓

   Context engineering curates useful instructions, evidence, tool outputs, and memory for the task.
2. Use embeddings as access controls for retrieved research

   Embeddings support similarity comparisons but do not establish authorization or replace access controls.
3. Place every document and tool result in each request

   Including everything can exceed context limits and obscure relevant information with unnecessary material.
4. Increase temperature to improve factual summaries

   Temperature affects sampling variability but cannot guarantee factuality or supply relevant evidence.

Context engineering selects and organizes relevant instructions, evidence, tool outputs, and memory for each request.

11. Evaluate the model using held-out data before deployment: Which TWO practices best support a sound feedback li

Select two. More than one option is correct — every correct one is ticked below.

1. Treat fluent responses as sufficient evidence of business value.

   Fluency does not establish factual accuracy, task quality, user outcomes, or financial value.
2. Use the test set repeatedly to select every new model version.

   Repeated test-set tuning contaminates the estimate of generalization and weakens confidence in final performance.
3. Evaluate the model using held-out data before deployment. ✓

   Held-out evaluation estimates generalization on data not repeatedly used for training or tuning.
4. Monitor production outcomes and use findings to guide improvement. ✓

   Monitoring reveals changing performance and business results, providing evidence for investigation and future improvements.
5. Retrain automatically whenever monitoring detects any performance change.

   A detected change requires investigation; automatic retraining can incorporate unchecked data or address the wrong cause.

Evaluate generalization before deployment and monitor real-world outcomes afterward, using evidence to guide responsible model improvement.

12. Use an agent that plans and selects approved tools: Which approach is best?

1. Use a foundation model only to generate a conversational reply.

   A conversational model can produce text, but text generation alone does not perform external system actions.
2. Use a text classifier that assigns each request to one response template.

   Classification can route requests, but it does not inherently plan multi-step actions or invoke tools.
3. Use an agent that plans and selects approved tools. ✓

   An agent can use a model to plan, choose tools, and adapt actions to the request.
4. Use a fixed workflow with predetermined steps for every request.

   A fixed workflow suits predictable sequences but cannot flexibly choose tools or steps for varying requests.

An agent is appropriate when the application must plan dynamically, select tools, and complete actions beyond conversational response generation.

Other AWS AI Practitioner domains

• Applications of Foundation Models — 390 questions →
• Fundamentals of AI and ML — 279 questions →
• Security, Compliance, and Governance for AI Solutions — 196 questions →
• Guidelines for Responsible AI — 195 questions →
• All 1396 AWS AI Practitioner questions →
• AWS AI Practitioner certification: requirements, cost and exam format →