Amazon Bedrock is a fully managed service from AWS that provides API-based access to foundation models from leading AI companies including Anthropic, AI21 Labs, Cohere, Meta, Mistral AI, Stability AI, and Amazon's own models. Rather than managing infrastructure, model hosting, or model training from scratch, developers can access state-of-the-art foundation models through a single unified API, customize them with proprietary data through fine-tuning or RAG, and build generative AI applications at scale with enterprise-grade security and privacy controls. Bedrock abstracts away the complexity of provisioning compute, managing model lifecycles, and securing inference endpoints, enabling teams to focus on application logic and business value. A key mental model: Bedrock is a model marketplace + inference runtime + customization toolkit — it's not a single model but an orchestration layer that lets you compare, customize, and operationalize many models using consistent tooling, with built-in safeguards like Guardrails to enforce compliance and responsible AI policies across all your generative AI workloads.
What This Cheat Sheet Covers
This topic spans 19 focused tables and 183 indexed concepts. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.
Table 1: Core Service Components
Amazon Bedrock's value comes from composing its managed building blocks — models, knowledge bases, agents, and safeguards — rather than using any single feature in isolation. Understanding each component and how they connect is the foundation for designing effective generative AI architectures on AWS.
| Component | Example | Description |
|---|---|---|
Claude 4.5, Llama 3, Nova 2 Lite, Nova 2 Sonic | • Pre-trained large language models, image generators, speech models, and embedding models from multiple providers available via unified API • each model optimized for different tasks | |
InvokeModel, InvokeModelWithResponseStream | • Synchronous or streaming inference calls to foundation models • supports on-demand, provisioned, batch, and tiered modes | |
Converse, ConverseStream | • Unified multi-turn conversation interface that works across all text models • provides consistent message format regardless of underlying model. | |
S3 → embeddings → OpenSearch | Managed RAG pipeline — automatically chunks, embeds, and indexes documents into vector stores for context-aware responses. | |
Agent with action groups + knowledge bases | Autonomous AI assistants that break down tasks, call APIs, query knowledge bases, and orchestrate multi-step workflows using foundation models. | |
Content filters (text + image), PII redaction | • Configurable safety and compliance controls applied to both inputs and outputs • blocks harmful content, enforces responsible AI policies across text and images | |
Runtime + Memory + Gateway + Browser + Identity | Enterprise agentic platform (GA) with 7 managed services — Runtime, Memory, Observability, Identity, Gateway, Browser, Code Interpreter — for deploying and operating production agents at scale with any framework. |