AWS Certified AI Practitioner (AIF-C01) - Practice Test 4

Amazon Bedrock Knowledge Bases enables fully managed RAG by automatically chunking documents, generating vector embeddings, and storing them in a vector database. At query time, it retrieves relevant content and injects it into the model prompt, grounding responses in actual company documents. Agents (A) orchestrate multi-step API workflows, not document-based Q&A. Guardrails (B) provide content filtering but no document retrieval. Model Evaluation (D) benchmarks model quality -- it doesn't retrieve documents at inference time. See more: Bedrock & Generative AI

Transfer learning reuses a model's previously learned visual representations (edges, shapes, textures) and adapts them to a new domain by retraining only the final classification layers on the new task. This requires far less training data and compute than training from scratch. Unsupervised clustering (A) has no class labels for target classification. RLHF (C) optimizes language model behavior from human preferences, not image classification. Synthetic data generation (D) can supplement training data but doesn't itself leverage the pre-trained model's learned features. See more: AI/ML Fundamentals

Model invocation logging captures the complete request and response payload for every model call, including the prompt, generated output, model ID, and timestamps. Logs are delivered to S3 and/or CloudWatch Logs for compliance and auditing. Guardrails (A) enforce content policies but don't capture full invocation records. Agents (C) orchestrate workflows but aren't a logging mechanism. CloudTrail (D) records API management calls (who called the API) but not the actual prompt and response content. See more: Bedrock & Generative AI

AI accountability requires three pillars: documentation (model cards, data lineage, design decisions), audit trails (logged records of when models were deployed, modified, and what decisions they produced), and clear ownership (named responsible individuals or teams for each system). Open-source models (A) provide code transparency but don't establish organizational accountability chains. On-premises deployment (C) is an infrastructure decision unrelated to accountability. Excluding PII (D) addresses privacy -- important, but separate from decision accountability. See more: AI Challenges & Responsibilities

At 15 invocations per day, On-Demand pricing is ideal because you pay only per token consumed with no minimum fees or commitments -- idle time costs nothing. Optimizing prompt token count further reduces each call's cost. Provisioned Throughput (A) requires monthly commitments for dedicated capacity, resulting in massive waste at 15 daily calls. Fine-tuning (B) adds training costs and requires Provisioned Throughput for inference. SageMaker dedicated endpoints (D) incur continuous hourly instance charges regardless of call volume. See more: Bedrock & Generative AI

Multimodal foundation models accept multiple input modalities -- images, text, and in some cases audio -- and generate text based on that combined context. Models like Claude 3 or Amazon Titan Multimodal can analyze product images and generate descriptive text. Text-only LLMs (A) cannot ingest image inputs. Image segmentation models (C) partition images into regions but don't generate descriptive text. OCR models (D) extract printed text from images but don't understand visual semantics or generate product descriptions. See more: Bedrock & Generative AI

Embedding models convert text into dense numerical vectors where semantic similarity corresponds to vector proximity. This enables cosine similarity or k-NN search to retrieve documents with similar meaning. Amazon Titan Embeddings and Cohere Embed are available via Amazon Bedrock for this purpose. Summarization models (A) condense documents but don't produce vectors for similarity search. Text generation models (B) produce text, not vectors optimized for semantic retrieval. Classification models (D) assign labels from a predefined set, not vector representations. See more: Bedrock & Generative AI

Amazon SageMaker Data Wrangler provides a visual, point-and-click interface for data preparation tasks: imputing or dropping missing values, detecting and capping outliers, standardizing date formats, and applying 300+ built-in transformations. It exports reusable processing pipelines to SageMaker. Athena (B) runs SQL queries for data exploration but doesn't offer interactive visual cleaning tools. Kinesis (C) handles real-time data streaming, not batch preparation. DMS (D) migrates databases between sources -- it is not a data preparation tool. See more: Amazon SageMaker

AWS Glue Data Catalog is a central metadata repository where Glue crawlers automatically scan data sources (S3, RDS, Redshift, and more), infer schemas, and register table definitions. Data scientists can then discover and query these assets from Athena, SageMaker, or EMR. S3 Inventory (B) produces reports of S3 object metadata but doesn't catalog cross-source datasets. Lake Formation (C) governs data lake access and builds on Glue Data Catalog but is not the cataloging service itself. QuickSight (D) is a business intelligence tool for visualization, not metadata cataloging. See more: AWS Cloud Computing

Amazon SageMaker JumpStart provides a model hub with 300+ pre-trained models that can be deployed to SageMaker endpoints with a single click or API call, handling all infrastructure provisioning automatically. EC2 with GPUs (B) requires manual environment setup, driver configuration, and model serving framework installation. Deep Learning AMIs (C) pre-install ML frameworks on EC2 but still require significant configuration. ECS (D) orchestrates containers but has no ML model hub or one-click deployment for pre-trained models. See more: Amazon SageMaker

Batch transform is designed for processing large offline datasets. It reads input from S3, processes all records using the model, writes outputs back to S3, and automatically releases infrastructure when done -- no persistent endpoint costs. Real-time inference (A) maintains always-on endpoints that would be idle most of the week, wasting money. Serverless inference (C) handles individual on-demand requests, not bulk scheduled processing. Edge deployment (D) runs models locally on IoT devices and is not applicable to centralized batch text processing. See more: Amazon SageMaker

Multimodal embeddings (like Amazon Titan Multimodal Embeddings) project both text and images into the same shared vector space. A text query and a semantically matching image will have similar vector representations, enabling searches that work across modalities. Word-level embeddings (A) operate only on text tokens. Sentence embeddings (C) represent text passages but cannot encode images. Knowledge graph embeddings (D) represent entities and relations in graph structures, not cross-modal visual-textual similarity. See more: Bedrock & Generative AI

AWS Glue is a fully serverless ETL service that runs Apache Spark jobs on automatically provisioned infrastructure. You write transformation scripts (or use visual tools), and Glue manages all compute resources, scaling based on data volume with no clusters to configure. EMR (B) provides managed Hadoop/Spark clusters but requires cluster sizing, configuration, and lifecycle management. Redshift Spectrum (C) queries S3 from within Redshift but is for SQL analytics, not ETL pipelines. Lambda (D) has a 15-minute execution limit and memory constraints, making it unsuitable for petabyte-scale data processing. See more: AWS Cloud Computing

SageMaker auto scaling monitors endpoint metrics (invocations per instance, CPU utilization) and automatically adjusts instance count based on configured policies. Target tracking policies add instances during spikes and remove them during quiet periods without manual intervention. Pipelines (A) orchestrate ML training and deployment workflows but don't respond to real-time inference traffic. Model Registry (C) manages model versions and approvals. Clarify monitoring (D) detects bias and data drift in production, not traffic-based capacity scaling. See more: Amazon SageMaker

A VPC gateway endpoint for S3 creates a private route between your VPC and S3 entirely within the AWS backbone network -- no internet traversal. Traffic is routed via a route table update at no additional cost. CloudFront (A) distributes content over the internet. Direct Connect (C) provides dedicated connectivity from on-premises networks to AWS, not between services within AWS. S3 Transfer Acceleration (D) routes uploads through CloudFront edge locations -- it traverses the internet. See more: AWS Security Services

BLEU (Bilingual Evaluation Understudy) measures translation quality by computing precision of matched n-grams (unigrams through 4-grams) between the generated translation and reference translations, with a brevity penalty. It is widely used as a reproducible, automated benchmark for MT systems. F1 score (B) balances precision and recall for classification tasks, not sequence generation. MAE (C) measures regression error between numerical predictions. Perplexity (D) evaluates how well a language model predicts a test set distribution, not translation accuracy against references. See more: AI/ML Fundamentals

Amazon Comprehend provides Named Entity Recognition (NER) that identifies and classifies entities including organizations, locations, people, dates, quantities, and monetary values. It can process documents in batch for high-volume workflows. Textract (B) extracts text and structured data from scanned images/PDFs but does not perform NER on the extracted content. Kendra (C) provides intelligent search over document collections but doesn't produce structured entity extraction. Translate (D) converts text between languages. See more: AWS Managed AI Services

Defense-in-depth against prompt injection requires: input validation (stripping or flagging patterns that attempt to override instructions), clear system prompt boundaries (separating system context from user input and explicitly instructing the model to ignore override attempts), and content filtering/guardrails (catching harmful outputs even if the prompt partially succeeds). Higher temperature (A) adds randomness but doesn't prevent injection. Smaller models (C) may behave differently but aren't inherently resistant to injection. No restrictions (D) eliminates all defenses. See more: AI Challenges & Responsibilities

Feature importance quantifies each input feature's contribution to the model's predictions. For gradient boosting models, this may use Gini importance or gain metrics. SHAP values (available via SageMaker Clarify) provide model-agnostic feature importance, showing each feature's directional and magnitude contribution to individual predictions. Feature engineering (B) is the process of creating or transforming features before training. Dimensionality reduction (C) compresses feature space rather than measuring individual contributions. Feature selection (D) chooses which features to use, often informed by importance analysis but is a distinct step. See more: AI/ML Fundamentals

Amazon Bedrock On-Demand pricing charges per input and output token with no upfront fees, no minimum consumption, and no long-term commitments -- perfectly suited for experimentation and variable workloads. Provisioned Throughput (A) requires 1-month or 6-month commitments for dedicated model units. EC2 Reserved Instances (C) apply to EC2 virtual machine capacity, not Bedrock API calls. Savings Plans (D) discount EC2, Lambda, and Fargate compute -- they don't apply to Bedrock token-based pricing. See more: Bedrock & Generative AI

SageMaker Clarify's model explainability feature computes SHAP (SHapley Additive exPlanations) values for individual predictions, showing each feature's positive or negative contribution to a specific outcome. For a single loan decision, it might show that debt-to-income ratio drove the score up while employment length drove it down. Pre-training bias detection (A) identifies class imbalances in training data before model training. Data validation (C) checks data schema and quality. Hyperparameter optimization (D) tunes training configuration -- none of these explain individual prediction outcomes. See more: Amazon SageMaker

Temperature controls the probability distribution over the next token. At temperature 0, the model applies argmax selection -- it always picks the single highest-probability token at each step, producing identical output for identical input. Temperature 1.0 (A) uses the raw distribution, introducing variability. max_tokens (C) limits response length but doesn't affect token selection randomness. Streaming (D) changes how the response is delivered but not the underlying token sampling strategy. See more: Prompt Engineering

k-NN search operates on vector representations, not raw text. Before any similarity search is possible, product descriptions must be converted into dense numerical vectors using an embedding model (such as Amazon Titan Embeddings via Bedrock). These vectors are then indexed in OpenSearch's k-NN index, and search queries are similarly converted at query time. Without embeddings, there are no vectors for k-NN to compute distances on. Dashboards (A), CloudWatch alarms (C), and S3 logging (D) are operational concerns that can be configured after the core search infrastructure is established. See more: Bedrock & Generative AI

Transparency is a core responsible AI principle. For medical AI, this requires explicit, persistent disclaimers that set appropriate expectations -- informing users that outputs are for general information only and do not replace professional diagnosis or treatment. This reduces the risk of harm from over-reliance. Unfiltered training data (A) risks including unreliable or harmful information. Autonomous diagnosis (C) is dangerous and potentially illegal without clinical validation. Disabling output filters (D) increases the risk of generating harmful medical misinformation. See more: AI Challenges & Responsibilities

SageMaker Asynchronous Inference supports payloads up to 1 GB and processing times up to 1 hour. Requests are queued and processed asynchronously, with results stored in S3. The endpoint can scale to zero during idle periods, eliminating costs between jobs. Real-time inference (A) has a 6 MB payload limit and 60-second processing timeout -- insufficient for 500 MB, 45-minute jobs. Serverless inference (B) is designed for lightweight, quick requests, not large long-running workloads. Multi-model endpoints (D) host multiple models on a single endpoint for cost efficiency, not large-payload handling. See more: Amazon SageMaker

Explicit formatting and length constraints in the prompt are the most reliable way to control output verbosity. Instructions like 'Respond in one sentence only' or 'Maximum 20 words' directly constrain the model's generation. Additional context (A) gives the model more to discuss, potentially worsening verbosity. Higher temperature (C) increases randomness but doesn't enforce brevity. A larger context window (D) increases how much input the model can process, not output length control. See more: Prompt Engineering

SageMaker Model Monitor continuously compares incoming production data distributions to a baseline captured from training data. It automatically detects statistical deviations (data drift), model quality degradation, bias drift, and feature attribution drift, emitting CloudWatch alerts when violations exceed configured thresholds. Debugger (B) monitors training jobs in real-time for gradient issues -- it is a training-time tool, not production monitoring. Clarify (C) provides static bias analysis and explainability reports but doesn't run continuous production drift monitoring. Experiments (D) tracks training runs for comparison. See more: Amazon SageMaker

Grounding connects an LLM's outputs to specific, external, authoritative documents. The most common implementation is RAG -- relevant document passages are retrieved and included in the prompt context, ensuring the model's response is anchored to real source material. This directly reduces hallucination. Fine-tuning (B) trains the model to better understand legal language but doesn't anchor responses to the specific contract at hand. Larger parameters (C) improve general capability but don't reduce hallucination in specific document contexts. A larger context window (D) allows longer input but doesn't inherently verify or anchor outputs. See more: Bedrock & Generative AI

Recall = TP / (TP + FN). Maximizing recall minimizes false negatives -- ensuring the model flags as many actual malicious connections as possible. Missing an attack (false negative) has a much higher cost than investigating a benign connection (false positive). Precision (A) minimizes false positives -- optimizing this would miss more attacks. Overall accuracy (C) doesn't separately penalize false negatives, which matters most here. Specificity (D) measures the true negative rate (correctly identifying benign traffic), which is the opposite of what needs optimization. See more: AI/ML Fundamentals

Amazon SageMaker Clarify provides pre-training and post-training bias metrics that measure disparities in model outcomes across defined protected groups. Metrics like Disparate Impact and Demographic Parity Difference quantify whether different groups receive systematically different predictions. Rekognition Face Analysis (A) processes images and video, not tabular candidate data. Comprehend Sentiment Analysis (C) analyzes text sentiment, not tabular prediction bias. Audit Manager (D) provides compliance evidence collection but no ML bias measurement. See more: Amazon SageMaker

TCO for build-vs-buy in ML must account for: (1) Upfront costs -- training a custom SageMaker model requires GPU instance hours, data labeling, and engineering time, while Comprehend charges per character with no training overhead; and (2) Ongoing costs -- custom models require continuous monitoring, retraining pipelines, and MLOps engineering to maintain performance, while Comprehend's managed service includes automatic updates. Programming language support (A), region (C), and console users (E) don't meaningfully factor into the cost comparison between these two approaches. See more: AI/ML Fundamentals

RAG with Knowledge Bases grounds the assistant's responses in actual company IT documentation. When a user asks about a specific error code, the system retrieves the relevant runbook section and provides it as context, producing company-specific answers. Higher temperature (A) increases randomness but doesn't provide access to company knowledge. Removing the system prompt (C) eliminates behavioral guidelines, worsening consistency. A larger model (D) has more general knowledge but still lacks company-specific documentation that was never in its training data. See more: Bedrock & Generative AI

Amazon Polly is AWS's text-to-speech service, converting written text to natural-sounding audio in dozens of languages and voices. It supports SSML for controlling pronunciation, speaking rate, and emphasis. Amazon Transcribe (A) performs the reverse operation -- converting spoken audio to text. Amazon Lex (C) builds conversational interfaces and chatbots, not audio synthesis. Amazon Translate (D) converts text between languages but does not generate audio output. See more: AWS Managed AI Services

Amazon Augmented AI (A2I) is built for routing ML predictions to human reviewers based on confidence thresholds. When a model's confidence falls below the configured threshold, A2I automatically creates a human review task with a customizable review interface and tracks review outcomes. SageMaker Ground Truth (B) is for labeling training datasets, not reviewing production model outputs. Mechanical Turk (C) is the workforce platform A2I can use internally, but using it standalone lacks automated confidence-based routing. Step Functions (D) orchestrates workflows but doesn't provide ML confidence-based human review interfaces. See more: Amazon SageMaker

SSE-C lets customers provide their own encryption keys with each S3 request. AWS uses the key to perform the encrypt/decrypt operation and then immediately discards it -- the key is never stored in AWS systems, and AWS has zero persistent access to the key material. SSE-S3 (A) and SSE-KMS (B) both store and manage key material within AWS infrastructure, meaning AWS technically has access to the keys. No encryption (D) provides no data protection at rest. See more: AWS Security Services

Collaborative filtering identifies users with similar behavioral patterns (viewing history, ratings) and recommends items that similar users enjoyed but the target user hasn't seen yet. It discovers latent preferences without requiring explicit feature engineering about movie genres. Amazon Personalize implements this at scale. Supervised classification (A) predicts from labeled features but requires predefined categories for every recommendation. Linear regression (C) predicts a continuous numerical value, not item recommendations. Anomaly detection (D) identifies unusual patterns rather than surfacing similarity-based recommendations. See more: AI/ML Fundamentals

SageMaker Model Registry provides model versioning (storing artifacts and associated metrics in a catalog), approval workflows (models transition through Pending Approval -> Approved -> Rejected states with audit trails), and deployment tracking (recording which version is deployed to which endpoint). Feature Store (B) manages ML feature data for training consistency, not model lifecycle governance. Experiments (C) tracks training runs and compares performance metrics but doesn't manage deployment approvals. CodePipeline (D) automates CI/CD pipelines but doesn't natively provide ML model versioning and approval workflows. See more: Amazon SageMaker

Amazon Bedrock requires Provisioned Throughput to serve inference from fine-tuned (custom) models. Fine-tuned models cannot be invoked via On-Demand pricing -- they require dedicated model units purchased under a 1-month or 6-month commitment. On-Demand (A) is only available for base foundation models, not customized variants. Batch inference (C) processes bulk jobs but fine-tuned models still require Provisioned Throughput to run. SageMaker endpoints (D) serve SageMaker-trained models -- models fine-tuned in Bedrock's customization workflow remain within the Bedrock service. See more: Bedrock & Generative AI

AI explainability requires making model behavior understandable to diverse audiences. Model cards provide structured documentation covering intended use, training data characteristics, evaluation results across groups, and known limitations -- consumable by non-technical stakeholders. SHAP-based feature importance explains how the model reaches decisions in terms of input variables. High complexity (A) typically reduces interpretability. Keeping details proprietary (C) is the opposite of explainability. Unsupervised models (D) don't inherently solve explainability -- any model type requires documentation and analysis. See more: AI Challenges & Responsibilities

Multimodal generation allows a single model invocation to produce outputs across multiple content types -- text and images. Amazon Bedrock models like Amazon Titan Image Generator create images from text prompts, while models like Claude can produce rich textual descriptions alongside image analysis. Text-only generation (A) produces only text -- images require a separate model invocation. Polly (C) synthesizes speech audio from text, not images. Rekognition (D) analyzes existing images, not generates new ones. See more: Bedrock & Generative AI

For consistent, high-volume, predictable workloads, Provisioned Throughput with a 6-month commitment provides dedicated model units at a significant discount (up to 50%) compared to On-Demand per-token pricing. When capacity is fully utilized across stable daily volume, the commitment discount delivers lower total spend. On-Demand (A) provides flexibility but at full per-token rates without volume discounts. Free Tier (C) is limited to initial trial usage, not production-scale. EC2 Spot Instances (D) discount EC2 compute but don't apply to Bedrock API pricing. See more: Bedrock & Generative AI

A significant gap between test set performance and production performance is the hallmark of distribution mismatch (covariate shift or train-serving skew). The test images don't represent the variety of lighting conditions, camera angles, or defect presentations seen in production. Too few parameters (A) would cause poor performance across both test and production (underfitting). An aggressive learning rate (C) causes training instability, not a deployment gap. The wrong loss function (D) would degrade training convergence, not produce a sudden production performance cliff. See more: AI/ML Fundamentals

Amazon Q in QuickSight provides a natural language querying interface where business users type questions in plain English and receive auto-generated visualizations and answers based on connected datasets, without any SQL knowledge. Athena (A) is a powerful SQL query engine but requires SQL proficiency. Redshift Query Editor (C) also requires SQL. Glue DataBrew (D) provides visual data preparation tools for data engineers, not natural language business querying. See more: Amazon Q

IP risk mitigation for generative AI requires: editorial review (humans check outputs for potential reproduction of source material), provenance transparency (using models where training data licensing is documented), and logging (maintaining records of generated content and timestamps for legal defensibility). No review (A) accepts unmitigated legal risk. Using competitor articles (C) constitutes direct copyright infringement as source material. Disabling safety filters (D) removes existing protections without addressing the IP problem. See more: AI Challenges & Responsibilities

AWS IAM enables resource-level access control for Amazon Bedrock. You create IAM policies specifying which principals (users, groups, roles) can invoke which specific model ARNs, providing the exact granularity required to allow access to some models but not others. Cognito (A) manages web and mobile application user authentication but doesn't directly govern which Bedrock model ARNs a user can invoke. WAF (C) protects web application endpoints from exploitation patterns, not API-level model authorization. Config (D) monitors resource configuration compliance but doesn't enforce Bedrock API authorization. See more: AWS Security Services

AWS Edge Locations are part of the CloudFront content delivery network (CDN). They cache copies of content (images, videos, web pages, APIs) geographically closer to end users to reduce latency. There are 400+ Edge Locations worldwide -- far more than Regions or Availability Zones. EC2 runs in Regions/AZs (A), database backups go to S3/other storage (B), and the console is a web application (D). See more: AWS Cloud Computing

The root account has unrestricted access to all AWS services and resources. Enabling MFA immediately after account creation is the single most important first security step, preventing unauthorized access even if the password is compromised. Creating an IAM admin user for daily use should follow. S3 buckets (A), EC2 instances (C), and Reserved Instances (D) are not security priorities. See more: AWS Cloud Computing

Amazon Q Developer's AI Code Companion integrates directly into VS Code and other IDEs to provide real-time code suggestions, code generation, and security scanning. It is specifically designed for AWS development with deep knowledge of AWS SDKs and services. CodePipeline (A) is for CI/CD. CodeGuru (B) does code reviews but not real-time suggestions. CloudFormation Linter (D) validates templates. See more: Amazon Q

AWS Chatbot deploys an AWS-aware bot into Slack and Microsoft Teams. With Amazon Q integration, engineers receive real-time alarm notifications and can ask troubleshooting questions directly in Slack, getting root cause analysis and remediation suggestions without leaving their chat tool. See more: Amazon Q

The principle of least privilege is a core AWS security best practice. It means granting only the minimum permissions a user or service needs to do their job -- nothing more. This reduces the blast radius if credentials are compromised. Defense in depth (A) is layered security. Zero trust (C) assumes no implicit trust. Separation of duties (D) divides responsibilities. See more: AWS Security Services