Hardest AWS Certified AI Practitioner questions this week

Unsupervised learning with clustering is designed to find structure in unlabeled data, which matches the scenario exactly. Supervised classification is wrong because it requires labeled training examples, and the tickets have no categories yet. Reinforcement learning applies to agents taking actions in an environment, not discovering groups in a static dataset.

Amazon Bedrock Knowledge Bases is the managed RAG feature that ingests from S3, handles chunking and embedding, stores vectors, and runs retrieval at query time. Agents orchestrate multi-step tool calls. Guardrails enforce content safety policies, and Provisioned Throughput is a pricing mode — none of these handle document ingestion and retrieval end-to-end.

Amazon SageMaker Clarify detects bias across groups before and after training and produces SHAP-based feature attributions for explainability. Model Monitor watches deployed models for data and quality drift but does not analyze bias or explain predictions. Debugger targets training-time issues, and Trusted Advisor is an account best-practices tool.

Recall measures the share of actual fraud cases the model catches, so optimizing it minimizes missed fraud — the costly error here. Precision is about how many flagged cases are truly fraud, which matters when false alarms are expensive. Accuracy is misleading on imbalanced fraud data because a model that predicts non-fraud for everything can still score high.

Foundation models are trained on broad data and generalize across many tasks through prompting, which is why one model can summarize, draft, and answer questions. Training a separate traditional model per task is exactly the overhead foundation models are designed to remove. Logistic regression and rule-based pipelines do not generalize to open-ended generative tasks.

Lower temperature reduces randomness in token sampling, producing more consistent and deterministic outputs — ideal for summarizing legal content. Higher temperature adds creative variation, which is the opposite of what this use case needs. Context window size affects how much text the model can read, not how deterministic its output is.

Vector embeddings map text into a continuous space where semantically similar phrases sit near each other, which is what enables matching "reset" with "recovery." Keyword indexes and regex dictionaries rely on exact token overlap and miss paraphrases. One-hot encodings carry no semantic similarity at all.