Amazon Bedrock
Generative AI model API
Call foundation models (Anthropic Claude, Amazon Nova, Meta Llama, Mistral…) through a single unified API, with no servers to manage; includes Knowledge Bases (RAG), Agents, Guardrails, and fine-tuning.
- Cost$$ Medium
- Difficulty●●●●●
- Costs money when idle?No
- Serverless
- In examsCLF · SAA
Think of it asLike an AI buffet: pick any dish (model) you want, AWS runs the kitchen, and you pay only for what you eat.
When to use it
✓ Use it when
Building generative AI apps (chatbots, summarization, RAG) without hosting models yourself.
✗ Avoid it when
You need to train your own model from scratch or control the entire infrastructure → SageMaker AI.
Common uses
- Q&A chatbot over internal documents (RAG)
- Text summarization and classification
- Agents that call APIs to perform tasks
What does it cost?
- Pricing model: On-demand is charged per input/output token (or per image, per second of video); Provisioned Throughput is billed hourly for committed throughput.
Provisioned Throughput and custom models are billed hourly even with no calls; long, repeated prompts consume many input tokens.
When the question says generative AI through an API, no servers to manage, choice of models, choose Bedrock; for answers based on frequently updated internal documents, Knowledge Bases (RAG) rather than fine-tuning; for filtering harmful content, Guardrails; for training your own model, SageMaker AI.
Frequently asked questions
How is Bedrock different from SageMaker AI?
Bedrock lets you call ready-made foundation models through an API to build generative AI applications, with no servers to manage or models to train. SageMaker AI is a full ML platform for preparing data and training and deploying your own models. For chatbots, summarization and RAG, use Bedrock; to build your own ML model, use SageMaker AI.
Is my data sent to Bedrock used to train models?
No. AWS commits that customer prompts and data are not used to train the base foundation models and are not shared with model providers. Data is encrypted, and you can reach Bedrock through PrivateLink so traffic stays off the Internet. Fine-tuning produces a custom model copy private to your account.
How do Knowledge Bases, fine-tuning and Guardrails differ?
Knowledge Bases use RAG: relevant documents are retrieved from your data and added to the prompt, suited to answers based on fresh, frequently changing data. Fine-tuning trains a model further to change its style or specialize in a task. Guardrails filter inputs and outputs according to your safety policies.
