Amazon AIF-C01: AWS AI/ML Services — Study Guide
Part of the AWS AI Practitioner AIF-C01 — Study Guide. Practice with verified answers in the Amazon exam hub, or take timed practice tests on ExamRoll.io.
Foundation models, Amazon Bedrock, and inference patterns
Foundation models hosted through Amazon Bedrock provide a fast path to generative capabilities using multiple model families, including high-throughput and instruction-tuned options such as Q and Nova. Choose between in‑context learning and customization: few‑shot prompts are cheap to iterate but sensitive to prompt injection and unstable across distribution shifts; parameter-efficient fine‑tuning or full customization reduces per‑inference cost and latency at the expense of data governance and retraining complexity. Inference cost and latency are determined primarily by model size and the number of input and output tokens processed, so design prompts to be concise, cache frequent responses, and prefer smaller specialized models for high-volume endpoints. Common practitioner traps include treating LLM outputs as authoritative without grounding, failing to control sensitive data fed into prompts, and not instrumenting prompt templates and versions. Decision criteria should weigh latency, token cost, privacy, and model ownership: use Bedrock’s private connectivity and KMS encryption when regulatory constraints demand data isolation, or choose SageMaker-hosted models when you need full control over training artifacts and explainability tools. Implement input sanitization, system-level instructions that are immutable, and layered safety checks to reduce prompt-based attacks and leakage of system prompts.
SageMaker for training, deployment, and model governance
Amazon SageMaker remains the primary managed path when full training control, explainability, and production-grade MLOps are required. Use SageMaker Training and distributed training instances for large datasets, SageMaker Studio for end‑to‑end development, and the Model Registry for versioning. For time series forecasting across many related SKUs, SageMaker DeepAR provides probabilistic forecasts and works best when you have many similar series and covariates. SageMaker Clarify helps detect bias and compute fairness metrics, while SageMaker Model Monitor tracks data drift and prediction quality in production. Common traps include leaving Studio notebooks with wide S3 permissions, under‑sizing instance types for distributed training, and skipping dataset provenance. Decision criteria between Bedrock and SageMaker often revolve around customization needs, regulatory controls, and explainability: pick Bedrock to rapidly prototype with high‑quality FMs and SageMaker when you require feature-level tracing, SHAP explanations, bias audits, or private, no‑internet execution environments using VPC endpoints and S3 VPC gateways.
Retrieval, multimodal ingestion, and grounding strategies
Reliable, factual outputs require grounding unstructured corpora with Retrieval Augmented Generation. Ingest documents using Textract for scanned forms and OCR, Transcribe and Polly for speech pipelines, Rekognition for image analysis and safety filters, and Comprehend for entity extraction and PII detection. Store vector embeddings in an OpenSearch or specialized vector DB; use Amazon Kendra when you need managed semantic search with rich connectors and relevance tuning. Architect chatbots with Bedrock for generation and Kendra or an embeddings store for retrieval, while Amazon Lex remains useful for intent routing and slot filling in deterministic flows. Agents for Bedrock are useful when orchestrating tool use, external APIs, or multi-step reasoning but require robust credentials and tool verifiers to avoid unsafe actions. Practitioner pitfalls include failing to paginate large corpora, ignoring retrieval latency impacts on user experience, and not versioning the retrieval index; decisions should balance freshness, index update cadence, and the cost of vector storage versus the risk of hallucination.
Security, privacy, explainability, and evaluation
Security and governance follow a shared responsibility model: AWS manages the service infrastructure, but the customer must classify data, manage IAM roles, encrypt data with KMS, and control network access via VPC endpoints and PrivateLink. For sensitive or regulated workloads, run SageMaker in a no‑internet VPC with S3 gateway endpoints and use Bedrock private networking options. Implement automated PII redaction before training using Comprehend and Textract pre‑processing, and consider differential privacy or data minimization when datasets contain confidential information. Explainability and bias detection require tools such as SageMaker Clarify, SHAP explainers, and human review workflows via Amazon Augmented AI for edge cases. Monitor models with Model Monitor metrics like feature drift, prediction drift, and custom factuality checks for RAG pipelines. Common mistakes are treating off‑the‑shelf model safety as sufficient, failing to log prompts and responses for audit, and not retaining validation datasets with controlled access; governance decisions should include deletion policies and incident response steps for model extraction or data leakage.
Practical Problem: Use-Case Scenario
Scenario: Nimbus Retail operates global brick-and-mortar stores and maintains sales and product image data in Amazon S3. Their AWS AI environment includes SageMaker Studio, Amazon Bedrock access, Amazon Kendra, OpenSearch Service, and Amazon QuickSight for BI.
Challenge: Nimbus needs an LLM-powered assistant to generate contextual product descriptions and localized translations while ensuring factual accuracy, protecting customer and supplier confidential data, and producing weekly top‑seller charts for store managers.
Recommended Approach:
- Use Amazon Textract to extract any structured attributes from product documents and Rekognition to validate and tag product images, storing cleansed assets in an encrypted S3 bucket with KMS and strict IAM roles.
- Build an embeddings index in OpenSearch or a managed vector store populated by embeddings from a Bedrock encoder; use Kendra for semantic retrieval of policy and spec documents.
- Implement a RAG pipeline: Bedrock for generation with retrieval context, apply a factuality filter using Comprehend and custom verification queries against the indexed corpus, and log all prompts and responses to CloudTrail and an audit S3 bucket.
- Produce weekly charts with Amazon QuickSight fed by aggregated sales from Athena queries over the S3 data lake, and surface model‑generated descriptions into a SageMaker endpoint for UX consumption with human review via A2I for high‑impact edits.
Rationale: This approach grounds generative outputs with retrieval and document verification to reduce hallucinations, secures data in encrypted S3 with least‑privilege access, and separates concerns by using managed services for indexing, generation, and BI to balance speed, governance, and accuracy.
← Generative AI Concepts · All domains · Responsible AI →
Practice these questions → · Timed practice on ExamRoll.io →
Pass the whole exam — not just this question
You found this answer. Get every verified question and explanation in one place, and save hours of prep. Free to start.
Pass your exam →