AWS Certified AI Practitioner 1-10

Amazon OpenSearch Service provides vector search (k-NN) capabilities, allowing you to store and search embeddings—numeric representations of text or images—as a vector database.
In RAG, a semantic similarity search is performed against these stored embeddings to retrieve relevant documents, which are then passed to the foundation model to improve the accuracy of its responses.
Memorizing the connection “storage location for embeddings = vector database = Amazon OpenSearch Service” is the shortcut to earning points here.
Ground Truth handles data labeling, Transcribe converts speech to text, and Textract extracts text and data from documents; none of them serve the role of storing vectors.
k-NN search – Amazon OpenSearch Service

The F1 score is a metric for evaluating the performance of a classification model, and it takes both precision and recall into account.
Precision represents “the proportion of items predicted as positive that were actually positive,” while recall represents “the proportion of actual positives that were correctly detected.”
The F1 score is the harmonic mean of precision and recall, expressing the balance between the two as a single value, which makes it especially useful when the data is skewed (class imbalance).
Speed, cost, and energy efficiency are different aspects of performance and are unrelated to the definition of the F1 score.
Evaluate the performance of models – Amazon Bedrock

Amazon Bedrock Guardrails is a feature that provides configurable safeguards for generative AI applications.
The reason it is correct is that it can detect and filter harmful content (hate, insults, sexual, violence, and so on) and remove PII (personally identifiable information) for both input prompts and model responses.
Agents is a feature that integrates with external systems to automatically carry out tasks, inference APIs are the inference APIs used to invoke models, and custom models are used for fine-tuning with your own data; none of them are intended for filtering harmful content itself.
Detect and filter harmful content by using Amazon Bedrock Guardrails

Amazon SageMaker AI is a fully managed service that lets you build, train, and deploy ML models in an integrated way.
Because you can repeatedly retrain the model with the latest data, it best fits the requirement of continuously improving accuracy.
Personalize specializes in recommendations and is not well suited to flexibly and iteratively training an arbitrary demand-forecasting model, CloudWatch is for monitoring and observability, and Rekognition is a service for image and video analysis; none of them fit the goal of improving the accuracy of a food-waste prediction model.
What is Amazon SageMaker AI?

Automatically forming groups (clusters) based on similarity from data that has no correct labels corresponds to clustering in unsupervised learning.
When you see the keywords “without labels, automatically grouping based on feature similarity,” you can judge it to be unsupervised learning.
Supervised learning learns prediction and classification from labeled data, reinforcement learning learns behavior through trial and error and rewards, and semi-supervised learning combines a small amount of labeled data with a large amount of unlabeled data; none of them are the focus of this requirement.
What is unsupervised learning? – AWS

LLM-as-a-judge is an evaluation method that uses another large language model (LLM) as a “judge” to automatically score subjective aspects of the output such as helpfulness, correctness, and tone.
It fits exactly the requirement of comparing multiple models on subjective quality dimensions in an automated and reproducible way without human involvement.
String matching checks for exact matches, and ROUGE is a metric that measures the word overlap between a summary and a reference text; neither is well suited to subjective evaluation such as tone.
RAG is not an evaluation method but a generation mechanism that retrieves external knowledge to reinforce the response.
Evaluate the performance of models – Amazon Bedrock

A regression model is a technique used to predict continuous values (numbers).
A task that predicts a continuous number such as “future temperature” is a classic use case for regression.
Suggesting a genre from a list is classification or recommendation, clustering movies is unsupervised learning, and generating an image is a task for generative models; none of them are prediction of continuous values.
Remembering the contrast “guessing a category = classification” and “guessing a number = regression” makes it easier to distinguish them.
What is machine learning (ML)? – AWS

Amazon Bedrock is a fully managed service that lets you use multiple foundation models (FMs) via API.
For the (generative AI) task of generating natural text such as hotel descriptions in a consistent writing style, Amazon Bedrock, which provides foundation models, is the best fit.
Comprehend is for analysis such as sentiment analysis and entity extraction from text, Personalize is for recommendations, and Rekognition is a service for image and video analysis; none of them are primarily intended for generating text in a consistent writing style.
What is Amazon Bedrock?

To make a model that reflects diversity, it is important to adjust the balance so that the representation of each group (class) in the training data is not skewed.
By balancing the representation of classes, you can suppress bias toward specific groups and get closer to a fair model that reflects diversity.
A region-specific dataset lacks global diversity, and oversampling the majority or deleting minority records both worsen the bias and undermine the reflection of diversity, so they are inappropriate.
Fairness, model explainability, and bias detection with SageMaker Clarify

Prompt injection is an attack in which malicious input overrides or diverts the original instructions configured in the system, causing unintended behavior.
Input that tries to hijack the original instructions, such as “Ignore the previous instructions…,” is a classic example of prompt injection.
Data poisoning is an attack that contaminates the training data, and a model inversion attack is an attack that infers the training data from the model’s output; neither matches this scenario.
Jailbreaking is close in that it aims to circumvent restrictions, but this example, which overrides existing instructions through input, is classified as prompt injection.
Detect and filter harmful content by using Amazon Bedrock Guardrails