MODULE 05 · BEDROCK

Amazon Bedrock

A 10-minute primer for your trainees. What Bedrock is, the three-layer stack, the key building blocks (Guardrails, Knowledge Bases, Agents), and how it sits underneath Amazon Quick.

~10 min Managed FMs Guardrails Knowledge Bases Agents

What is Amazon Bedrock?

A fully managed AWS service for building and scaling generative AI applications with foundation models — via a single API, without managing any infrastructure.

The 30-second answer

Bedrock is the developer's entry point to generative AI on AWS. Call any top foundation model — Anthropic Claude, Amazon Nova, Meta Llama, Mistral, Amazon Titan — through one API, add safety Guardrails, ground answers in your own documents with Knowledge Bases, or wire up an Agent that takes multi-step actions. Everything sits inside your AWS security perimeter.

Where Quick is the no-code surface for business users, Bedrock is the API + toolkit for engineering teams building durable AI products. They share the same underlying models.

The three-layer AWS AI stack

Applications · AI-powered products
TOP OF STACK

The apps end users touch. Amazon Quick, Amazon Q, and your own custom applications — chatbots, agents, in-product AI features — all live here.

Amazon Bedrock · managed FM tools
DEVELOPER API

The service that exposes foundation models to developers, plus the batteries — Guardrails for safety, Knowledge Bases for RAG, Agents for multi-step reasoning, Model Evaluation for choosing the right one.

Foundation models · the core intelligence
MODELS

The pre-trained models themselves — Claude, Nova, Llama, Mistral, Titan Embeddings — pre-integrated into Bedrock, no separate downloads or deployment.

Models on Bedrock · pick the right one for the job

Anthropic
Claude Sonnet / Opus
Reasoning, long-form drafts, coding assistance
Amazon
Nova Micro · Lite · Pro
Fast and cost-efficient across text tasks
Meta
Llama 3
Open-weight option for many general tasks
Mistral
Mistral / Mixtral
Efficient generation, multilingual support
Amazon
Titan Embeddings
Vector encoding for RAG and semantic search
Amazon
Nova Canvas / Reel
Image and short-form video generation
The trainer's one-liner: "You don't pick a model on Bedrock — you pick the right model for each task. Cheap and fast for classification, larger and slower for reasoning. One API, many models."

The building blocks trainees should know

You don't need to teach every Bedrock feature. Four names, one clear job each — that's the level.

Model Invocation
The core API — send a prompt, get a completion. One InvokeModel or Converse call across every model.
FOUNDATION
Guardrails
Configurable safety filters that run before + after model calls. Block denied topics, redact PII, filter hate/violence, ground answers against your allowed data.
SAFETY
Knowledge Bases
Managed RAG. Point at documents in S3, Bedrock chunks + embeds + indexes them. Ask a question, get an answer grounded in your data with citations.
GROUNDING
Agents
Multi-step reasoning that can call your APIs (via action groups) or MCP tools. Plan → act → observe → repeat, all within the Bedrock managed loop.
ACTION

How they fit together

A production Bedrock app usually stacks these: Agent orchestrates the reasoning, calling a foundation model via Model Invocation, wrapped in Guardrails for safety, grounded in a Knowledge Base for the domain data. That's four names covering ~80% of real-world Bedrock architectures.

Things Bedrock also has (know they exist)

FeatureWhat it does
Model EvaluationCompare models on your own tasks with automatic and human judges — pick the right model with data, not vibes.
Prompt ManagementVersion, test, and share prompts as a first-class artefact — not scattered in code.
FlowsVisual builder for GenAI workflows — chain a KB retrieval into a Guardrail into a model call.
Data AutomationExtract structured data (fields, tables) from unstructured docs and images at scale.
MarketplaceAccess to third-party specialised models beyond the built-in providers.
Provisioned ThroughputReserve dedicated model capacity for consistent latency at production scale.
Don't teach these in your Quick session. Just show trainees they exist so the phrase "the model I need is on Bedrock" doesn't sound scary. A curious engineer will read the Skill Builder course on their own time.

Quick vs Bedrock · when to reach for which

The one comparison your trainees will actually need. Same models underneath, very different surfaces.

Amazon Quick — reach for it when…

  • The user is a business team, not an engineer
  • You want a working prototype today, not next sprint
  • The output is text, a small app, or a research brief
  • Grounding = "point at this folder", not "design a vector store"
  • Everything happens inside quick.aws.com or the AWS Console

Amazon Bedrock — reach for it when…

  • Building a customer-facing product with strict SLAs
  • Need fine-grained control over prompts, models, retrieval
  • Need to integrate AI into an existing app or backend
  • Compliance / governance / audit-log requirements demand it
  • You're a developer with an IDE and a git repo

Same job · which tool wins

SituationToolWhy
"Give our merchant-support team a policy Q&A assistant"Quick (Space + agent)Point at the folder, share the link. Ships this afternoon.
"Add AI-powered dispute triage to our contact-centre app"Bedrock (Agent + KB)Custom integration inside an existing SLA-bound product.
"Prototype an ideation workflow for the founder team"Quick (chat)Faster to iterate the prompt than to write code around it.
"Ship a mobile-app feature that summarises long PDFs"Bedrock (Model Invocation + Guardrails)Native API call from your app, versioned prompts, safety controls.
"Explain 'what can AI do for us?' to a non-technical exec"QuickYou can show it live in 2 minutes. Bedrock needs a code demo.
"Fine-grained control over vector-store chunking for 5M docs"Bedrock Knowledge BasesConfigure indexes, embedding models, retrieval scoring. Quick abstracts this away.

Quick sits on top of Bedrock

When trainees ask about the relationship, this is the answer: the chat model behind Quick is a Bedrock foundation model. The Space grounding uses managed RAG concepts identical to Bedrock Knowledge Bases. The chat agents parallel Bedrock Agents. Quick is Amazon's opinionated, no-code surface built on the same rails engineers use for custom builds.

Which means: anything a trainee prototypes in Quick, an engineer can productionise on Bedrock later — same model family, same safety patterns, same architecture. That's a powerful message to leave with the room.

Round-trip closing line: "You've seen Quick chat, research, apps, and grounded agents. Bedrock is where those ideas graduate when you need a real product. Different tools, same intelligence, one AWS."
End the session here

This tab is your natural closer. If you've done the previous four modules right, trainees leave knowing: (1) what Quick does, (2) how to run an ideation demo, (3) where to find free content to keep learning, (4) what Kiro is if they build software, (5) what Bedrock is if they want to go deeper. Five landing pads. That's the whole training day.