AIF-C01 question patterns

A managed AI service or Amazon Bedrock called through an API.

Training and hosting your own model needs data, skills and time that the question rules out.

Prompt engineering first, then RAG; fine-tuning only if those fall short.

Prompting needs no training. Fine-tuning and continued pre-training add compute, data prep and hosting cost.

Don’t use ML; use deterministic rules or code.

ML models predict and can be wrong. They can’t guarantee an exact result.

Choose a smaller or efficient model and avoid training from scratch.

Energy use scales with model size and training effort.

Amazon Textract.

Rekognition finds text in photos; Textract understands document structure.

Amazon Rekognition.

It is the pre-trained computer vision API.

Amazon Comprehend.

It is the pre-trained NLP service for analysing text.

Amazon Lex.

Lex handles dialogue flow and language understanding for bots.

Amazon Kendra.

Intelligent enterprise search with connectors.

Amazon Personalize.

Managed recommender built on your user and item data.

SageMaker Canvas.

It is the no-code ML interface.

Amazon Q Business.

It connects to company data and respects existing permissions.

Amazon Q Developer.

It is the AI coding assistant for IDEs, CLI and console.

Amazon Bedrock.

JumpStart needs you to choose instances and host endpoints.

SageMaker JumpStart.

You choose the instance, container and hosting setup.

Provisioned throughput.

Custom models in Bedrock are served only with provisioned throughput; on-demand is for variable, low traffic.

RAG with Bedrock Knowledge Bases.

It retrieves fresh content at question time without retraining and reduces hallucinations.

Aurora or RDS for PostgreSQL with pgvector.

It stores embeddings alongside existing relational data.

Amazon Neptune.

Graph database with vector support.

Fine-tuning (instruction tuning).

Examples of prompt and ideal response change model weights to fit the style or task.

Continued pre-training.

It extends general knowledge with unlabeled data using the pre-training objective.

Agents for Amazon Bedrock with action groups.

Agents plan steps and call APIs and knowledge bases.

Lower the temperature.

Low temperature makes token choice less random.

Raise the temperature (or top-p).

More randomness gives more diverse output.

Few-shot prompting.

Zero-shot gives none; single-shot gives one.

Chain-of-thought prompting.

Step-by-step reasoning improves accuracy on complex problems.

Prompt injection; defend with Guardrails prompt-attack filters, least privilege and untrusted-input handling.

Jailbreaking targets safety rules; injection hijacks the task.

ROUGE.

It measures overlap with reference summaries.

BLEU.

It scores n-gram precision against reference translations.

BERTScore.

It uses embeddings so paraphrases score well.

Optimise recall.

Recall counts false negatives; precision counts false positives.

Look at precision, recall and F1, not accuracy.

A model that always says “not fraud” scores 99% accuracy yet catches nothing.

Guardrails for Amazon Bedrock.

It applies content, topic, word and sensitive-information filters on input and output.

Guardrails contextual grounding check (and RAG).

It flags responses not supported by the supplied sources.

SageMaker Clarify.

It reports bias metrics and feature attributions.

SageMaker Model Monitor.

It watches live inference data for drift and bias drift.

Amazon Augmented AI (A2I).

It builds human-in-the-loop review workflows.

SageMaker Model Cards.

They are the standard transparency record.

Biased or unrepresentative data; use subgroup analysis and add diverse, balanced data.

More data of the same skewed kind doesn’t fix it.

AWS PrivateLink interface VPC endpoint.

It keeps traffic on the AWS network.

Amazon Macie.

It discovers sensitive data in S3. Guardrails handles PII at inference time.

IAM policies scoped to specific model ARNs.

Least privilege applies to model invocation like any other action.

AWS Artifact.

Audit Manager gathers your own evidence; Artifact supplies AWS’s reports.

AWS Audit Manager.

It maps controls to evidence, including a generative AI framework.

AWS CloudTrail (plus Bedrock invocation logging for prompts).

CloudTrail records API activity.

AWS Config.

It records configuration and evaluates it against rules.

Data lineage and cataloguing (Glue Data Catalog, SageMaker lineage, Model Cards).

Provenance records support trust and audits.

Generative AI Security Scoping Matrix (five scopes).

The more you own, the more security responsibility you carry.