Module 8 · AWS AI Services

Amazon Bedrock Explained

Amazon Bedrock is a managed service for building generative AI applications with foundation models and related capabilities.

Written and reviewed by Mahendra Akula Knowledge Bytes

Article available nowVideo lesson coming later

This reading lesson is complete and can be studied independently. The video will be added here when it is published.

Core concept

What Amazon Bedrock is

Amazon Bedrock is a fully managed AWS service for building generative AI applications with foundation models. Instead of provisioning and operating the infrastructure that hosts a model, an application invokes managed model APIs. Teams can experiment with models, compare results, add private knowledge, apply safeguards, evaluate quality, and move the selected design into an application.

A foundation model is trained on broad data and can support many downstream tasks. Depending on the selected model and modality, a Bedrock application can generate or transform text, work with images, create embeddings, summarize information, answer questions, or support a larger agentic workflow. The model is only one part of a production system: prompts, data, retrieval, permissions, evaluation, monitoring, and user experience determine whether the complete application is useful.

Simple definition: Amazon Bedrock gives applications managed access to foundation models and the supporting services needed to build governed generative AI solutions on AWS.

Architecture diagram

How a Bedrock application works

1UserQuestion or task

2ApplicationPrompt and business logic

3Amazon BedrockManaged inference API

4Foundation modelGenerate or transform

5ResponseValidate and present

Knowledge BasesRetrieve approved business context and citations.

GuardrailsEvaluate inputs and outputs against configured policies.

EvaluationCompare quality, safety, latency, and cost.

The user sends a request. The application receives a question, document, image, instruction, or task.
The application prepares context. It creates the prompt, applies business rules, and may retrieve information from a knowledge base.
Bedrock invokes the selected model. The application identifies the model and inference settings through the supported API.
The model produces an output. The output might be text, an image, an embedding, or another supported modality.
The application validates the result. It can apply safeguards, check citations, log metrics, request human review, and present the final result.

The major Bedrock capabilities

Model choice and inference

Use models available through the Bedrock model catalog and invoke them through managed APIs. Model support varies by provider, feature, modality, and AWS Region, so selection must be based on the actual workload.

Knowledge Bases and RAG

Retrieve information from connected sources and add it to the prompt. This grounds an answer in organizational information and can return citations that help users verify the response.

Guardrails

Evaluate user inputs and model responses using configured policies such as content filters, denied topics, sensitive-information filters, and contextual-grounding checks.

Model evaluation

Compare candidate models and application behavior against criteria such as answer quality, task success, safety, latency, throughput, and cost before choosing a production design.

Model customization

When prompting and retrieval are insufficient, supported models can be adapted with techniques such as fine-tuning. Customization requires suitable data, evaluation, and additional operational decisions.

Agents and orchestration

Agentic systems can interpret a request, use knowledge, invoke approved tools, and work through multiple steps. Permissions and human approval must match the impact of the actions.

Practical example

Build an internal employee-support assistant

Imagine a company wants employees to ask questions about leave, expenses, and benefits. A general model may produce a plausible answer from broad knowledge, but company policies are private and change over time. The solution therefore needs retrieval, access control, citations, and an escalation path.

1. Governed sourcesApproved policy documents are connected to a Bedrock Knowledge Base. Document access must match employee permissions.

2. RetrievalAn employee question is converted into a search request and the most relevant policy passages are retrieved.

3. Grounded generationThe application sends the question and retrieved passages to a selected foundation model through Bedrock.

4. Safe responseGuardrails evaluate the interaction, the answer includes citations, and uncertain or high-impact cases are escalated to HR.

5. EvaluationThe team measures groundedness, correctness, retrieval quality, latency, cost, and employee feedback over time.

Amazon Bedrock or Amazon SageMaker AI?

The services can be complementary, but they start from different needs. Select from the requirement rather than memorizing one service as universally better.

Decision area	Amazon Bedrock	Amazon SageMaker AI
Typical starting point	Build a generative AI application with managed foundation-model access.	Build, train, tune, deploy, or govern broader machine-learning workloads.
Infrastructure focus	Managed model inference and GenAI application capabilities.	Greater control over the ML lifecycle and model-development environment.
Common examples	RAG assistants, content generation, summarization, embeddings, and GenAI workflows.	Custom predictive models, training jobs, feature engineering, endpoints, and MLOps.
Exam clue	The scenario emphasizes foundation models or managed generative AI capabilities.	The scenario emphasizes building or managing a custom ML model and lifecycle.

Security and responsible AI

Production considerations

Identity and permissions

Use least-privilege IAM permissions for model invocation, data sources, knowledge bases, tools, logs, and administrative actions.

Data protection

Classify information, use encryption and secure transport, understand retention settings, and prevent sensitive data from entering uncontrolled fields or logs.

Grounding and verification

Use approved sources, expose citations where appropriate, measure groundedness, and keep human review for consequential or uncertain outputs.

Monitoring and improvement

Track quality, safety, latency, errors, cost, and changing user behavior. Re-evaluate after model, prompt, data, or policy changes.

AWS Certified AI Practitioner

What to remember for the exam

Recognize these scenario signals

Managed foundation-model access: consider Amazon Bedrock.
Answers based on private documents: consider RAG and Bedrock Knowledge Bases.
Input and output safeguards: consider Bedrock Guardrails as one layer of the responsible-AI design.
Choosing between models: compare task quality, modality, latency, cost, context needs, and supported features.
Custom ML lifecycle: distinguish the scenario from the broader build-and-train capabilities of SageMaker AI.

Key takeaways

Choose model
Invoke through managed API
Add knowledge and safeguards
Evaluate application

Check your understanding

Can you explain this concept in two sentences without using jargon?
Can you identify the input, process, output, and validation step in the example?
Can you name one suitable use case and one case where another approach is better?
Which risk or limitation should a responsible implementation address?

Research references

Public sources and further reading

This lesson is original educational writing informed by the public references below. Use the sources to explore definitions, technical details, and current AWS exam objectives.