Amazon Bedrock

Amazon Bedrock: Complete Guide for the AIF-C01 Exam

Why Amazon Bedrock Is Important

Amazon Bedrock is one of the most heavily tested topics on the AWS Certified AI Practitioner (AIF-C01) exam. It represents AWS's flagship service for building generative AI applications and is central to understanding how organizations can leverage foundation models (FMs) without managing complex infrastructure. As generative AI becomes a core pillar of modern cloud strategy, Amazon Bedrock serves as the primary gateway through which AWS customers access, customize, and deploy large language models and other foundation models in a secure, scalable, and responsible manner.

What Is Amazon Bedrock?

Amazon Bedrock is a fully managed service that provides access to a choice of high-performing foundation models (FMs) from leading AI companies — including Amazon (Titan), Anthropic (Claude), Meta (Llama), Cohere, Mistral AI, Stability AI, and AI21 Labs — through a single, unified API. It allows developers and organizations to build and scale generative AI applications without needing to provision, manage, or maintain any infrastructure.

Key characteristics of Amazon Bedrock include:

• Serverless: There is no infrastructure to manage. You simply call the API and pay for what you use.
• Multi-model access: Choose from a wide variety of foundation models from multiple providers, all accessible through a single service.
• Customization: You can fine-tune models with your own data or use Retrieval Augmented Generation (RAG) through Knowledge Bases for Amazon Bedrock.
• Security and Privacy: Your data is not used to train the base foundation models. Data remains within your AWS account and is encrypted in transit and at rest.
• Enterprise-ready: Integrates with AWS services like IAM, CloudWatch, CloudTrail, VPC, and PrivateLink for governance, monitoring, and security.

How Amazon Bedrock Works

Amazon Bedrock operates through several core components and capabilities:

1. Foundation Model Access
You select a foundation model from the Bedrock model catalog. Each model has different strengths — for example, Claude excels at nuanced text generation and analysis, Titan offers a versatile Amazon-built option, Stable Diffusion specializes in image generation, and Llama provides an open-source alternative. You invoke these models through API calls (InvokeModel or InvokeModelWithResponseStream for streaming).

2. Model Customization
Bedrock supports two key methods of customization:
• Fine-tuning: You provide labeled training data to adapt a foundation model to your specific domain or task. This creates a custom model that is private to your account.
• Continued Pre-training: You provide unlabeled data to teach the model domain-specific knowledge without task-specific formatting.
• Both methods use your data securely — training data is encrypted and never shared with model providers.

3. Knowledge Bases for Amazon Bedrock (RAG)
This feature enables Retrieval Augmented Generation, which allows models to reference your proprietary data sources when generating responses. You connect data sources (e.g., Amazon S3), and Bedrock automatically chunks, embeds, and stores this data in a vector database (such as Amazon OpenSearch Serverless, Pinecone, or Amazon Aurora). At inference time, the model retrieves relevant context before generating a response, significantly reducing hallucinations and improving accuracy.

4. Agents for Amazon Bedrock
Agents allow foundation models to execute multi-step tasks by dynamically invoking APIs and interacting with external systems. You define the agent's instructions and the action groups (backed by AWS Lambda functions) it can call. The agent uses the FM's reasoning capabilities to break down a user request, determine which actions to take, execute them, and return results. This is ideal for building AI assistants that can perform real-world tasks like booking appointments, querying databases, or managing workflows.

5. Guardrails for Amazon Bedrock
Guardrails allow you to implement responsible AI safeguards across your generative AI applications. You can configure:
• Content filters: Block harmful content across categories like hate, insults, sexual content, violence, and misconduct.
• Denied topics: Prevent the model from engaging with specific topics (e.g., giving investment advice).
• Word filters: Block specific words or phrases.
• Sensitive information filters (PII): Detect and redact or block personally identifiable information.
• Contextual grounding checks: Detect hallucinations by checking whether responses are grounded in the provided source material.
Guardrails can be applied to any model in Bedrock and work on both input (prompts) and output (responses).

6. Model Evaluation
Amazon Bedrock provides built-in tools to evaluate model performance using:
• Automatic evaluation: Uses built-in metrics like accuracy, robustness, and toxicity.
• Human evaluation: Allows human reviewers to assess model outputs using custom or built-in criteria, leveraging your own workforce or an AWS-managed workforce.

7. Provisioned Throughput
For production workloads requiring consistent performance, you can purchase Provisioned Throughput, which guarantees a specific level of model inference capacity. This is necessary to deploy fine-tuned or custom models.

8. Amazon Bedrock Pricing
Bedrock offers two primary pricing models:
• On-Demand: Pay per input/output token (for text models) or per image (for image models). No commitment required.
• Provisioned Throughput: Pay for a reserved number of model units with a commitment term (1-month or 6-month).

Key Integration Points

• Amazon S3: Store training data, knowledge base source documents, and model artifacts.
• AWS Lambda: Execute actions for Bedrock Agents.
• Amazon OpenSearch Serverless: Serve as the vector store for Knowledge Bases.
• AWS IAM: Control access to models, features, and resources.
• AWS CloudTrail: Log all Bedrock API calls for auditing.
• Amazon CloudWatch: Monitor invocation metrics and set alarms.
• AWS PrivateLink: Access Bedrock privately without traversing the public internet.

Amazon Bedrock vs. Amazon SageMaker

This is a common exam comparison:
• Amazon Bedrock: Fully managed, serverless access to pre-built foundation models. Ideal for teams that want to use, customize, and deploy FMs with minimal ML expertise and no infrastructure management.
• Amazon SageMaker: A comprehensive ML platform for building, training, and deploying custom machine learning models from scratch. Ideal for data scientists and ML engineers who need full control over the ML lifecycle.
• Amazon SageMaker JumpStart: Provides access to pre-trained foundation models (similar to Bedrock) but within the SageMaker environment, requiring more infrastructure management.

The exam often tests whether Bedrock or SageMaker is the better fit for a given scenario. If the question emphasizes ease of use, no infrastructure, serverless, or using existing foundation models, the answer is typically Bedrock.

Exam Tips: Answering Questions on Amazon Bedrock

Here are critical strategies and knowledge points for tackling Amazon Bedrock questions on the AIF-C01 exam:

Tip 1: Know the Core Value Proposition
Amazon Bedrock is serverless and fully managed. If a question asks about accessing foundation models without managing infrastructure, Bedrock is almost always the answer.

Tip 2: Understand Data Privacy
A frequently tested concept: Your data is never used to train the base foundation models in Bedrock. Custom model training data stays within your account. This is a key selling point for enterprise adoption and regulatory compliance.

Tip 3: RAG = Knowledge Bases for Amazon Bedrock
When a question describes a scenario where a model needs to reference company-specific or up-to-date information to reduce hallucinations, the answer involves Knowledge Bases for Amazon Bedrock (RAG). Remember that RAG does not modify the model — it augments the prompt with retrieved context.

Tip 4: Multi-Step Tasks = Agents for Amazon Bedrock
If the scenario involves an AI assistant that must perform actions (call APIs, query databases, execute workflows), think Agents for Amazon Bedrock.

Tip 5: Responsible AI = Guardrails for Amazon Bedrock
Any question about content filtering, blocking harmful outputs, preventing PII leakage, or enforcing topic restrictions should lead you to Guardrails for Amazon Bedrock. Remember that guardrails can detect hallucinations through contextual grounding checks.

Tip 6: Fine-Tuning vs. RAG
Know when to use each approach:
• Fine-tuning: When you need to change the model's style, tone, or teach it a specific task format. Requires labeled training data.
• RAG: When you need the model to access current, dynamic, or proprietary information. Does not modify the model weights.
• The exam may present scenarios where both could work — look for keywords like up-to-date information (RAG) or specific output format/style (fine-tuning).

Tip 7: Model Selection Matters
Be familiar with the types of models available:
• Text generation: Claude (Anthropic), Titan Text (Amazon), Llama (Meta), Command (Cohere), Mistral
• Image generation: Stable Diffusion (Stability AI), Titan Image Generator (Amazon)
• Embeddings: Titan Embeddings (Amazon), Cohere Embed
• If a question asks about generating embeddings for a vector database, think Amazon Titan Embeddings.

Tip 8: Bedrock vs. SageMaker Distinction
This will appear on the exam. Remember: Bedrock = use existing FMs with minimal effort; SageMaker = build and train custom models with full control. If the question mentions custom algorithms, data scientists, training from scratch, or full ML lifecycle, think SageMaker. If it mentions foundation models, serverless, quick deployment, or generative AI applications, think Bedrock.

Tip 9: Model Evaluation
Know that Bedrock provides both automatic and human evaluation capabilities. Automatic evaluation uses metrics like BERTScore, toxicity, and accuracy. Human evaluation can involve your own team or AWS-managed workers.

Tip 10: Security and Compliance
Bedrock encrypts data at rest and in transit. It supports VPC endpoints via AWS PrivateLink for private connectivity. All API calls are logged in CloudTrail. These details are important for questions about secure and compliant generative AI deployments.

Tip 11: Watch for Distractor Services
The exam may include options like Amazon Comprehend, Amazon Lex, Amazon Rekognition, or Amazon Polly. These are purpose-built AI services, not generative AI foundation model services. Don't confuse them with Bedrock. If the question is about generative AI with foundation models, Bedrock is the answer.

Tip 12: PartyRock
AWS PartyRock is a playground built on Amazon Bedrock that lets you experiment with foundation models without an AWS account. If a question mentions experimenting or prototyping with generative AI in a no-code environment, PartyRock could be relevant.

Summary

Amazon Bedrock is the cornerstone of AWS's generative AI strategy and a critical topic for the AIF-C01 exam. Remember its key features: serverless access to multiple foundation models, customization through fine-tuning and RAG, responsible AI through Guardrails, task automation through Agents, and enterprise-grade security. Understand when to use Bedrock versus SageMaker, and always consider the specific capabilities (Knowledge Bases, Agents, Guardrails) that match the scenario described in each exam question.