Amazon Bedrock
Fully managed access to foundation models through one API.
❓ What is it?
A serverless service that exposes foundation models from Amazon (Nova, Titan), Anthropic (Claude), Meta (Llama), Mistral, and others through a single API — no infrastructure to provision and no model hosting to manage.
💡 Why does it exist?
Training a foundation model costs millions; most teams only need to USE one. Bedrock removes the undifferentiated heavy lifting — capacity, scaling, model updates — so builders focus on prompts, retrieval, and application logic while keeping data private (prompts are not used to retrain base models).
⏱️ When should you use it?
Choose Bedrock when you want generative AI (chat, summarisation, RAG, agents, image generation) without managing GPUs or model weights. Prefer SageMaker instead when you must train or fully control a custom model.
🗺️ Where does it fit?
It sits behind your application tier: a Lambda function or backend service calls the Bedrock runtime API. Knowledge Bases connect it to your documents in S3; Agents connect it to your business APIs; Guardrails wrap every call.
🔌 How do you integrate it?
Enable model access in the console, then call InvokeModel / Converse via the AWS SDK. Add Knowledge Bases for RAG (Bedrock handles chunking, embedding, and vector storage), Agents for tool-use, and fine-tuning or continued pre-training jobs for customisation with your own data.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- Know the customisation ladder in order of cost/effort: prompt engineering → RAG (Knowledge Bases) → fine-tuning → continued pre-training.
- On-Demand pricing bills per input/output token; Provisioned Throughput buys dedicated capacity and is required for customised models.