1. What is Amazon Bedrock?

Amazon Bedrock is a fully managed service that provides access to high-performing foundation models (FMs) from leading AI companies through a single API. It is the easiest way to build and scale generative AI applications on AWS.

Core Concept
Bedrock = serverless GenAI. Choose a foundation model (Claude, Titan, Llama, Mistral, etc.), call the API with a prompt, and get a response. No GPUs to provision, no model weights to download, no infrastructure to manage. Your data is NOT used to train the base models. Customize with RAG, fine-tuning, or agents.

2. Foundation Model Providers

Model Selection
Choose based on task: Claude = best reasoning + safety. Titan Embeddings = best for RAG vector embeddings. Llama = fine-tuning flexibility (open weights). Mistral = cost-effective for simpler tasks. Stable Diffusion = image generation. The exam tests your ability to pick the right model for a scenario.

3. How Bedrock Works

Bedrock API Flow:

1. Your Application
     |
2. Call Bedrock API (InvokeModel)
   - Model ID: anthropic.claude-3-sonnet
   - Prompt: "Summarize this document: ..."
   - Parameters: temperature=0.3, max_tokens=1000
     |
3. Bedrock routes to the selected Foundation Model
   - Runs inference on managed infrastructure
   - NO GPU management, NO model hosting
     |
4. Response returned to your application
   - Generated text, token usage, stop reason
     |
5. Your data stays in YOUR account
   - NOT used to train the base model
   - Encrypted in transit and at rest

4. Bedrock API Modes

5. Model Customization


Fine-Tuning

  1. Train a foundation model further on YOUR labeled data
  2. Model learns your domain-specific patterns, terminology, and style
  3. Requires: labeled training data in JSONL format, stored in S3
  4. Creates a Custom Model version (stored in your account)
  5. Supported models: Titan Text, Llama, Cohere Command (not all models support fine-tuning)
  6. Use for: domain-specific language, custom output format, specialized classification


Continued Pre-Training

  1. Train on YOUR unlabeled data to teach the model new domain knowledge
  2. Does NOT need labeled examples — just a large corpus of domain text
  3. Example: feed the model thousands of internal documents, legal texts, and medical literature
  4. Use for: domain adaptation without labeled data (teach new vocabulary, concepts)

6. Bedrock Guardrails

Guardrails help you implement safeguards for your generative AI applications, controlling model inputs and outputs.

  1. Content filters: block harmful, violent, sexual, hate speech, or insults (configurable thresholds)
  2. Denied topics: define topics the model should refuse to discuss (e.g., competitor products, political opinions)
  3. Word filters: block specific words or phrases in inputs/outputs
  4. PII filters: detect and redact/block personally identifiable information (SSN, email, phone, credit card)
  5. Contextual grounding: check if model responses are grounded in the provided context (reduce hallucination)
  6. Apply to: any Bedrock model invocation, Knowledge Bases, Agents
  7. Use for: responsible AI, compliance, safety, brand protection

7. Model Evaluation

  1. Compare model outputs for your specific use case before choosing
  2. Automatic evaluation: use built-in metrics (accuracy, robustness, toxicity)
  3. Human evaluation: set up human reviewers to rate model quality
  4. Use for: selecting the best model, validating the custom model quality

8. Bedrock Pricing

Important Warning
Bedrock does NOT use your data to train base models. Your prompts, responses, and custom training data stay in your account and are encrypted. This is critical for enterprise adoption and a common exam question about Bedrock data privacy.
Exam Tip
Bedrock FMs: "Serverless GenAI API" = Bedrock. "Multiple model providers" = Bedrock (Claude, Titan, Llama, Mistral, etc.). "Fine-tune FM" = Bedrock Custom Models. "Continued pre-training" = domain adaptation without labels. "Block harmful content" = Guardrails. "Stream responses" = InvokeModelWithResponseStream. "Multi-turn chat" = Converse API. "Batch processing" = Batch Inference (50% cheaper). Data privacy: YOUR data is NOT used to train base models.