Amazon Bedrock — A Complete Beginner's Guide
Amazon Bedrock lets you use powerful AI models through an API without building or managing any infrastructure. It is one of the most frequently tested topics on the AIF-C01 exam.
---
What Is Amazon Bedrock?
Bedrock is AWS's fully managed generative AI service. You choose from a catalog of foundation models (FMs) from multiple providers and call them via a standardized API.
Think of it like renting a fully equipped kitchen — you pick the menu (use case), choose the chef (model), and serve the food (output), without building the kitchen yourself.
Four core characteristics to memorize: Fully managed: no server setup or maintenance Data privacy: your data never leaves your AWS account Pay-per-use: no upfront commitment Unified API: same interface across all models
---
Available Foundation Models
| Provider | Model | Strength | |---|---|---| | Amazon | Titan | Multimodal, fine-tunable, cost-efficient | | Anthropic | Claude | Largest context window (200K), complex reasoning | | Meta | Llama | Large-scale text tasks | | Stability AI | Stable Diffusion | Image generation only |
---
Fine-Tuning — Customizing the Model
Three approaches, each with different costs and use cases:
| Method | Data Needed | Cost | Best For | |---|---|---|---| | Instruction-Based | Prompt-completion pairs | Lowest | Domain Q&A, chatbots | | Continued Pre-training | Unlabeled text (bulk) | Higher | Deep domain specialization | | Conversational | Multi-turn dialogue | Medium | Customer service bots |
---
RAG — Connecting Real-Time Information
RAG (Retrieval-Augmented Generation) lets the model search your documents at query time instead of memorizing everything during training.
When to use RAG vs fine-tuning:
| Situation | Recommended | |---|---| | Data changes frequently | RAG | | Fixed domain expertise | Fine-tuning | | Need fast document integration | RAG | | Adjusting tone or style | Fine-tuning |
---
Pricing Models
| Option | Discount | Best For | |---|---|---| | On-Demand | None | Unpredictable traffic | | Batch | Up to 50% | Overnight bulk processing | | Provisioned Throughput | Long-term discount | Stable, high-volume workloads |
Custom/fine-tuned models MUST use Provisioned Throughput — on-demand deployment is not available.
---
Exam Key Points
Bedrock = fully managed + data privacy guaranteed. Changing data → RAG; fixed domain knowledge → fine-tuning. Batch mode saves up to 50%. Custom models require Provisioned Throughput. Temperature, Top K, Top P do NOT affect cost or latency. Guardrails filter at input, processing, and output layers.