2.1.1

Question	Answer & Explanation
### Question 1 - Document Extraction: Managed AI vs. Foundation Models A logistics company needs to automate the extraction of text, tables, and form data from millions of scanned, multi-page shipping manifests. The manifests vary significantly in layout across different global vendors. The engineering team must implement a solution with the least development effort and lowest operational overhead. Which approach best meets these requirements? - A) Train a custom computer vision model using Amazon SageMaker AI to identify and extract text from the varied document layouts. - B) Use Amazon Textract to automatically extract text, tables, and form data from the scanned manifests. - C) Use a multimodal foundation model in Amazon Bedrock, passing the scanned document images as base64 encoded strings in the prompt to ask for the extracted text. - D) Use Amazon Comprehend to parse the text and extract key entities from the shipping manifests.	Correct Answer: B Correct Option Analysis: - B is Correct: Amazon Textract is a managed AWS AI service specifically pre-trained to extract text, tables, forms, and structured data from scanned documents. It requires no custom model training or infrastructure management, satisfying the least development effort and lowest operational overhead constraints. Distractor Analysis: - A is Incorrect: Training a custom computer vision model on SageMaker requires data labeling, model training, and hosting infrastructure, resulting in high development effort and operational overhead. - C is Incorrect: While multimodal FMs in Bedrock can analyze images, using them for high-volume OCR extraction of complex tables is less accurate, much slower (higher latency), and significantly more expensive than a dedicated service like Textract. - D is Incorrect: Amazon Comprehend is a natural language processing (NLP) service used for extracting entities and sentiment from text, not for performing OCR on scanned images. > Exam Tip / AWS Best Practice: For already-solved, standardized tasks like OCR, image recognition, or speech-to-text, always prefer a managed AWS AI service over training a custom model or using a Foundation Model.
### Question 2 - Customizing Model Behavior: Prompting vs. Fine-Tuning A customer service team is using an Amazon Bedrock foundation model to generate automated responses to customer inquiries. The model's responses are grammatically correct, but they frequently fail to use the company's specific internal terminology and preferred tone. The team needs to correct this behavior with the fastest time to deployment and minimal infrastructure management. Which approach should the team take first to address this issue? - A) Provision a custom model in Amazon Bedrock using Supervised Fine-Tuning on a dataset of historical customer service transcripts. - B) Increase the max_tokens parameter in the Bedrock API request to allow the model to generate longer responses. - C) Implement task-specific prompt engineering by adding clear instructions and a few-shot examples of the desired tone and terminology to the prompt. - D) Use Amazon Bedrock Knowledge Bases to ingest the company's style guide and retrieve it at inference time. (Select ONE)	Correct Answer: C Correct Option Analysis: - C is Correct: Task-specific prompt engineering (using instructions and few-shot examples) is the fastest and most operationally efficient way to adapt a foundation model's style, tone, or formatting. It changes behavior immediately without requiring data preparation, training jobs, or model versioning. Distractor Analysis: - A is Incorrect: Supervised fine-tuning requires preparing labeled datasets, running training jobs, and managing model versions. It should only be used as a last resort when prompt engineering fails to close the performance gap. - B is Incorrect: Increasing max_tokens only allows for longer outputs; it does not influence the semantic style, tone, or terminology used by the model. - D is Incorrect: While Bedrock Knowledge Bases (RAG) is useful for injecting external factual knowledge, enforcing a specific tone or style is a behavioral task better handled by prompt engineering rather than retrieving a style guide document. > Exam Tip / AWS Best Practice: Always follow the "Prompt First, Fine-Tune Second" rule. Fine-tuning should only be considered when evaluations consistently show that prompt engineering cannot achieve the required task performance.
### Question 3 - Knowledge Delivery: RAG vs. Customization An enterprise legal team is building an AI assistant to answer questions about internal corporate policies. The policy documents are updated frequently, and the legal team mandates that every answer generated by the assistant must explicitly cite the exact source document and section. The solution must reflect policy edits within minutes of publication. Which architectural approach meets these requirements? - A) Fine-tune a foundation model in Amazon Bedrock on the full text of the corporate policies so the knowledge is baked into the model weights. - B) Use Amazon Bedrock Knowledge Bases to ingest the policy documents into a vector store and retrieve relevant passages at query time. - C) Train a custom text-generation model using Amazon SageMaker AI on historical legal Q&A pairs. - D) Use Amazon Comprehend to extract key phrases from the policies and return them to the user as answers. (Select ONE)	Correct Answer: B Correct Option Analysis: - B is Correct: Amazon Bedrock Knowledge Bases (Retrieval-Augmented Generation) fetches relevant document passages at query time. Because the content lives in a vector store, updates reflect within minutes of re-ingestion. Furthermore, retrieved chunks carry source metadata, allowing the model to explicitly cite the source document and section. Distractor Analysis: - A is Incorrect: Fine-tuning writes knowledge into model weights. It cannot easily cite exact sources, and updating policies would require a new, time-consuming training job rather than reflecting changes in minutes. - C is Incorrect: Training a custom model from scratch on SageMaker AI requires massive compute, extensive labeled data, and does not natively support the citation of dynamic source documents. - D is Incorrect: Amazon Comprehend is an NLP service for entity extraction and sentiment analysis, not a generative AI solution capable of synthesizing natural language answers or managing document citations. > Exam Tip / AWS Best Practice: Use RAG when knowledge changes frequently or when source citations are required. Use fine-tuning when you need to change the model's behavior, tone, or format, not its factual knowledge.
### Question 4 - Forecasting: Trained ML vs. Foundation Models A regional grocery chain needs to generate weekly demand forecasts for 5,000 SKUs across 100 store locations. They have years of historical sales data, including features for seasonality and promotions. The finance department requires that the model's predictions be highly interpretable so they can understand the driving factors behind each forecast. Inference costs must remain low. Which modeling approach is the most appropriate? - A) Use a large language model in Amazon Bedrock, passing the historical tabular data as a CSV string in the prompt to ask for next week's forecast. - B) Train an interpretable supervised machine learning model, such as XGBoost, on the historical tabular data using Amazon SageMaker AI. - C) Fine-tune a multimodal foundation model on the historical sales data using Amazon Bedrock. - D) Write a fixed rule-based script that calculates a rolling 4-week average for each SKU. (Select ONE)	Correct Answer: B Correct Option Analysis: - B is Correct: The answer (patterns of seasonality and promotions) is already in the company's historical data. Training a traditional supervised ML model like XGBoost on SageMaker AI is designed precisely for tabular time-series forecasting. XGBoost is also highly interpretable (feature importance can be easily extracted) and runs with very low inference cost compared to foundation models. Distractor Analysis: - A is Incorrect: Foundation models (LLMs) are not designed for precise, large-scale numerical time-series forecasting. Passing massive tabular datasets into the context window is highly inefficient, extremely expensive, and fails the interpretability requirement. - C is Incorrect: Fine-tuning a foundation model on tabular sales data is an anti-pattern. FMs are for general language and creative tasks, not structured regression problems. - D is Incorrect: Fixed rules (rolling averages) cannot capture complex, dynamic demand patterns, seasonality, or the impact of promotions, resulting in poor forecast accuracy. > Exam Tip / AWS Best Practice: When the answer lives in your proprietary historical data, train a traditional ML model. When the capability is a general language task, use a foundation model.
### Question 5 - Fine-Tuning Strategies: Distillation Preparation A machine learning engineer wants to optimize an Amazon Bedrock foundation model for a specific coding assistant task. To reduce inference costs and latency, the engineer decides to use distillation to transfer the performance of a larger, more capable teacher model into a smaller, faster student model. Which data preparation step is specifically required for the distillation process in Amazon Bedrock? - A) Preparing a dataset of human-labeled, golden prompt-response pairs in JSONL format. - B) Preparing a dataset of prompt-response pairs where the responses are generated by the teacher model. - C) Preparing a reward function that scores model responses based on code execution success. - D) Preparing a dataset of image-text pairs to align multimodal capabilities. (Select ONE)	Correct Answer: B Correct Option Analysis: - B is Correct: In model distillation, the engineer prepares a set of prompts, and the managed service automatically generates the responses using the larger teacher model. The student model is then fine-tuned on these teacher-generated responses to mimic the larger model's behavior for that specific task. Distractor Analysis: - A is Incorrect: Human-labeled, golden prompt-response pairs are used for Supervised Fine-Tuning (SFT), not distillation. - C is Incorrect: A reward function that scores responses is used for Reinforcement Fine-Tuning, not distillation. - D is Incorrect: Image-text pairs are for multimodal fine-tuning, not specifically for the text/coding distillation process described. > Exam Tip / AWS Best Practice: Distillation uses a teacher model to generate the training labels for a student model, reducing the labeling effort while transferring task-specific capabilities to a cheaper, faster model.
### Question 6 - Model Selection: Latency vs. Cost Constraints A startup is building a real-time chatbot application using Amazon Bedrock. The application must provide responses to user queries with a maximum end-to-end latency of 1.5 seconds. The team is evaluating multiple foundation models and must balance output quality with cost. The queries are simple factual questions that do not require complex multi-step reasoning. Which foundation model selection strategy will best meet the latency and cost constraints? - A) Select the largest available parameter model (e.g., Anthropic Claude 3 Opus) to ensure the highest output quality and fewest hallucinations. - B) Select the smallest, fastest available model (e.g., Anthropic Claude 3 Haiku) that meets the baseline accuracy requirements during evaluation. - C) Select a mid-sized model (e.g., Anthropic Claude 3 Sonnet) and enable Provisioned Throughput to guarantee lower latency. - D) Select the largest model and increase the max_tokens parameter to cap the response length at 1.5 seconds. (Select ONE)	Correct Answer: B Correct Option Analysis: - B is Correct: For simple factual questions, a smaller, faster model (like Haiku) typically provides sufficient accuracy while offering significantly lower latency and lower cost per token compared to larger models. Selecting the smallest model that meets the quality bar is the standard best practice for real-time, cost-sensitive applications. Distractor Analysis: - A is Incorrect: The largest models have the highest latency and cost. Using them for simple queries violates both the 1.5-second latency constraint and the need to balance cost. - C is Incorrect: While Provisioned Throughput guarantees capacity, a mid-sized model still inherently has higher latency than a smaller model. If a smaller model meets the accuracy needs, it is the better choice for strict low-latency constraints. - D is Incorrect: max_tokens controls the length of the output (and thus the time to generate), but the initial time-to-first-token (TTFT) for a large model processing the prompt is still too slow to reliably meet a 1.5s requirement for simple tasks. > Exam Tip / AWS Best Practice: Model selection is a measurement of tradeoffs. For real-time applications with strict latency limits, evaluate smaller models first; only move to larger models if evaluations show a quality gap that prompting cannot close.
### Question 7 - Avoiding Catastrophic Forgetting A healthcare company is fine-tuning an Amazon Bedrock foundation model on a highly specialized dataset of internal medical notes. During testing, the engineering team notices that while the model excels at formatting the internal medical notes, it loses its ability to accurately answer general medical questions or write coherent emails, a capability the base model previously had. Which strategy should the team implement to mitigate this issue? - A) Decrease the number of epochs during the fine-tuning job to lower the training time. - B) Blend broader, general-domain data into the fine-tuning training mixture alongside the specialized medical notes. - C) Increase the temperature parameter during inference to encourage more diverse outputs. - D) Switch to a Retrieval-Augmented Generation (RAG) architecture using Amazon Bedrock Knowledge Bases. (Select ONE)	Correct Answer: B Correct Option Analysis: - B is Correct: The scenario describes "catastrophic forgetting," where a model trained hard on a narrow domain loses its general capabilities. The standard defense is to mix a broader, general-domain dataset into the specialized training data so the model retains its foundational language abilities while learning the new task. Distractor Analysis: - A is Incorrect: While decreasing epochs might slightly reduce overfitting, it will also reduce the model's ability to learn the specialized medical formatting. Blending data is the direct solution to retaining general capabilities. - C is Incorrect: Adjusting the temperature at inference time affects the randomness of token selection but does not restore lost semantic capabilities or knowledge in the model weights. - D is Incorrect: Switching to RAG is a valid architectural choice for injecting external knowledge, but the question asks how to mitigate the forgetting during the fine-tuning process itself. > Exam Tip / AWS Best Practice: Catastrophic forgetting occurs when a model is fine-tuned too narrowly. Blending general data into the training mixture helps the model retain its broad capabilities while adapting to the specific domain.
### Question 8 - Choosing AWS AI Services for Image Analysis A retail company wants to build a moderation pipeline to automatically detect inappropriate content, identify specific brand logos, and categorize scenes in user-uploaded images on their mobile application. The team has no in-house computer vision expertise and must launch the feature within two weeks with zero model training required. Which AWS service should the engineering team use? - A) Amazon Comprehend - B) Amazon Rekognition - C) Amazon SageMaker JumpStart - D) Amazon Bedrock (Select ONE)	Correct Answer: B Correct Option Analysis: - B is Correct: Amazon Rekognition is a fully managed computer vision service that requires zero machine learning expertise or model training. It natively supports content moderation, object/scene detection, and custom labels (brand logos) out of the box, meeting the strict two-week timeline and zero-training constraint. Distractor Analysis: - A is Incorrect: Amazon Comprehend is a natural language processing (NLP) service for text analysis, not a computer vision service for image analysis. - C is Incorrect: Amazon SageMaker JumpStart provides pre-trained models, but deploying and integrating them still requires ML engineering, endpoint management, and operational overhead, violating the "zero model training/expertise" constraint. - D is Incorrect: While Amazon Bedrock has multimodal models, using an FM to do basic content moderation and scene categorization is overkill, slower, and more expensive than a purpose-built managed service like Rekognition. > Exam Tip / AWS Best Practice: Do not use a foundation model or custom training for already-solved tasks. For standard image and video analysis, Amazon Rekognition is the most operationally efficient and cost-effective choice.
### Question 9 - Foundation Model Customization: Reinforcement Fine-Tuning A financial services firm is using an Amazon Bedrock foundation model to generate investment risk summaries. The firm wants to improve the model's ability to format the summaries according to strict compliance rules (e.g., always including a specific disclaimer and avoiding speculative language). The team has domain experts who can define a set of rules to score the model's outputs based on compliance, rather than providing perfect text examples. Which customization strategy in Amazon Bedrock is best suited for this use case? - A) Supervised Fine-Tuning using a dataset of perfectly written compliance summaries. - B) Continual Pre-Training on a large corpus of financial documents. - C) Reinforcement Fine-Tuning using a reward function that scores model responses based on compliance rules. - D) Provisioned Throughput to ensure the model processes compliance rules faster. (Select ONE)	Correct Answer: C Correct Option Analysis: - C is Correct: Reinforcement Fine-Tuning allows users to improve a model's reasoning and formatting by providing a reward function (or grader) that scores the model's outputs based on specific criteria (in this case, compliance rules). This is ideal when domain experts can define what a "good" answer looks like via rules rather than writing hundreds of perfect examples. Distractor Analysis: - A is Incorrect: Supervised Fine-Tuning requires a dataset of perfectly written prompt-response pairs (labeled input/output). The prompt states experts will define rules to score outputs, not write perfect summaries. - B is Incorrect: Continual Pre-Training is used to inject new domain knowledge (e.g., a new language or internal jargon) into a model, not to enforce strict output formatting or compliance rules. - D is Incorrect: Provisioned Throughput is an account-level billing/capacity feature; it does not alter the model's behavior or enforce compliance rules. > Exam Tip / AWS Best Practice: Use Supervised Fine-Tuning when you have labeled input/output pairs. Use Reinforcement Fine-Tuning when you can define a reward function to score the model's outputs based on business rules.
### Question 10 - Pre-trained Models vs Foundation Models: SageMaker JumpStart A data science team is developing an anomaly detection system for IoT sensor data. They have a proprietary dataset of time-series readings and want to use a pre-trained model as a starting point, but they need full control over the training algorithm, hyperparameters, and the underlying inference endpoints to optimize for extreme low-latency edge deployments. They do not need natural language generation capabilities. Which AWS service provides the most appropriate set of pre-trained models for this use case? - A) Amazon Bedrock - B) Amazon Textract - C) Amazon SageMaker JumpStart - D) Amazon Bedrock Knowledge Bases (Select ONE)	Correct Answer: C Correct Option Analysis: - C is Correct: Amazon SageMaker JumpStart provides access to pre-trained models and built-in algorithms specifically for traditional machine learning tasks (like XGBoost for anomaly detection, time-series forecasting, etc.). It allows data scientists full control over the training pipeline, hyperparameters, and endpoint configurations, which is ideal for proprietary tabular/IoT data and edge deployments. Distractor Analysis: - A is Incorrect: Amazon Bedrock provides access to Foundation Models (LLMs and multimodal models) via API. It does not provide traditional ML models for tabular anomaly detection, nor does it offer the granular control over training algorithms and edge deployment infrastructure required here. - B is Incorrect: Amazon Textract is a managed AI service for OCR; it is not a machine learning service for anomaly detection on IoT data. - D is Incorrect: Bedrock Knowledge Bases is a managed RAG service for document retrieval, entirely unsuitable for numerical anomaly detection on IoT time-series data. > Exam Tip / AWS Best Practice: For traditional ML tasks (tabular data, time-series, anomaly detection) requiring custom training pipelines and endpoint control, use SageMaker JumpStart. For generative AI and language tasks, use Amazon Bedrock.