How to Use Amazon Bedrock: Complete Beginner's Guide (2026)

By the end of this guide, you will have made your first API call to a foundation model on Amazon Bedrock, connected a document store through Knowledge Bases for retrieval-augmented generation, and configured Guardrails to filter harmful content before it reaches your users. Every step includes a verification checkpoint so you know it worked before moving on.

Bedrock is not a model. It is a fully managed, serverless platform that gives you API access to over 100 foundation models from providers including Anthropic (Claude), Meta (Llama), Mistral AI, Amazon (Nova), and others. You never provision a GPU, manage an endpoint, or download model weights. If your infrastructure already runs on AWS, Bedrock turns generative AI into a configuration choice rather than an infrastructure project. For a deeper look at what the platform offers, see our What Is Amazon Bedrock breakdown.

What You Need Before Starting

Amazon Bedrock runs inside the AWS Management Console. Unlike consumer AI chatbots, you need an AWS account with billing enabled and appropriate IAM permissions before you can make your first request. The service itself has no minimum spend or upfront commitment on the Standard tier. You pay per token (or per image, per second of audio) only when you use it.

Step 1: Accessing the Bedrock Console

The Bedrock console is your command center for model management, testing, and monitoring. Here is how to get there.

1. Sign in to the AWS Management Console with your IAM credentials.
2. In the search bar at the top, type Bedrock and select Amazon Bedrock from the results.
3. Confirm your region in the top-right dropdown. Select us-east-1 (N. Virginia) or us-west-2 (Oregon) for the broadest model catalog.
4. You will land on the Bedrock dashboard showing model providers, usage metrics, and quick-start links.

The left navigation panel organizes Bedrock's features: Model catalog (browse and enable models), Playgrounds (test models interactively), Knowledge bases, Agents, Guardrails, Flows (visual pipeline builder), and Custom models (fine-tuning and distillation).

Step 2: Requesting Model Access

Bedrock does not enable all models by default. Each model provider requires you to request access before you can invoke it. This is a one-time step per model, per region.

1. In the Bedrock console, click Model catalog in the left sidebar.
2. Browse available providers: Amazon (Nova family), Anthropic (Claude), Meta (Llama), Mistral AI, AI21 Labs, Cohere, DeepSeek, Stability AI, and others.
3. Click on the model you want to enable. For this guide, start with Amazon Nova Micro (cheapest at $0.035/1M input tokens) and Anthropic Claude 3.5 Haiku (fast, cost-effective at $0.25/1M input tokens).
4. Click Request model access. Most Amazon and Meta models are approved instantly. Anthropic models may take a few minutes.
5. Repeat for any additional models you want to test.

The Bedrock Marketplace (accessible from the model catalog) adds another 100+ models beyond the core catalog, including specialized models for code generation, medical text, and domain-specific tasks.

Step 3: Making Your First API Call

Bedrock offers two paths for your first interaction: the Console Playground (no code required) and the AWS SDK (programmatic access). Start with the Playground to verify everything works, then move to code.

Option A: Console Playground (Fastest)

1. In the left sidebar, click Playgrounds > Chat.
2. Select a model from the dropdown. Choose Amazon Nova Micro for a quick, cheap test.
3. Type a prompt: "Explain Amazon Bedrock in three sentences for a developer who has never used it."
4. Click Run. The response appears in the output panel within seconds.
5. Experiment with the parameters panel on the right: Temperature (0 = deterministic, 1 = creative), Max tokens (response length), and Top P (nucleus sampling).

Option B: AWS CLI

With the AWS CLI v2 installed and configured, run this command to invoke the Converse API (AWS's unified format that works across all Bedrock models):

aws bedrock-runtime converse --model-id amazon.nova-micro-v1:0 --messages '[{"role":"user","content":[{"text":"What is Amazon Bedrock?"}]}]' --region us-east-1

The response includes the model's output text, token usage counts, and a stop reason. You pay only for the tokens consumed by this call.

Option C: Python SDK (boto3)

Bedrock supports five API formats: Converse API (unified, recommended), Invoke API (provider-native), Messages API (Anthropic-native for Claude), Chat Completions API (OpenAI-compatible), and Responses API (newer OpenAI patterns). The Converse API is the best starting point because it works with every model without format changes.

Step 4: Setting Up Knowledge Bases (Managed RAG)

Knowledge Bases give your models access to your own data through retrieval-augmented generation (RAG). Instead of the model relying only on its training data, it searches your documents and grounds its answers in your specific content. Every response includes source attribution so you can trace answers back to the original document.

1. Navigate to Knowledge bases in the left sidebar and click Create knowledge base.
2. Name your knowledge base and select an IAM role (Bedrock can create one for you).
3. Choose a data source. Bedrock supports S3 buckets, Confluence, Salesforce, SharePoint, and web crawlers.
4. For a quick start, create an S3 bucket, upload 5-10 PDF or text documents, and select that bucket as your data source.
5. Configure the chunking strategy. The default (fixed-size chunks) works for most use cases. Advanced options include semantic chunking and hierarchical chunking.
6. Select a vector store. Bedrock can create a managed OpenSearch Serverless collection automatically, or you can connect Aurora PostgreSQL, MongoDB Atlas, Pinecone, or Redis.
7. Click Create and wait for the sync to complete. Bedrock handles chunking, embedding, and vector storage automatically.
8. Once synced, test it in the Knowledge Base Playground: ask a question about your documents and verify the response cites specific sources.

Step 5: Building a Bedrock Agent

Bedrock Agents handle multi-step task automation with tool use, code interpretation, and memory retention across sessions. An agent can break down a complex request into sub-tasks, call external APIs, query Knowledge Bases, and synthesize a final answer.

1. Navigate to Agents in the left sidebar and click Create Agent.
2. Name your agent and write clear instructions describing its role and behavior. Example: "You are a customer support agent for [company]. Answer questions about our products using the product knowledge base. If the customer needs to place an order, collect their email and product selection."
3. Select a foundation model. Claude 3.5 Haiku ($0.25/1M input tokens) is a strong default for agents that need fast responses. Nova Pro ($0.80/1M input tokens) works well for agents that need multimodal understanding.
4. Attach your Knowledge Base from Step 4 so the agent can search your documents.
5. Define action groups: these are the tools your agent can use. Each action group maps to an API schema (OpenAPI format) or a Lambda function. For a basic agent, start with just the Knowledge Base. Add action groups later as you need external integrations.
6. Enable Code Interpreter if your agent needs to run calculations, generate charts, or process data files.
7. Click Prepare to compile the agent, then test it in the Agent Playground.

For complex workflows, multi-agent collaboration lets you chain specialized agents together. AgentCore, which reached general availability in late 2025, extends this further with its own runtime, gateway, memory, identity management, and policy engine. AgentCore works with any framework (LangChain, CrewAI, custom) and any model, not just Bedrock models.

Step 6: Configuring Guardrails

Guardrails add a safety layer between your users and the model. AWS reports that Bedrock Guardrails block 88% of harmful content (AWS documentation). Six configurable safety policies let you control exactly what gets through.

1. Navigate to Guardrails in the left sidebar and click Create guardrail.
2. Name your guardrail and add a description.
3. Configure Content filters: set sensitivity thresholds (None, Low, Medium, High) for hate speech, insults, sexual content, violence, and misconduct. Start with "High" filtering across all categories.
4. Add Denied topics: describe topics your application should refuse to discuss. Example: "Investment advice," "Medical diagnosis," "Competitor product recommendations."
5. Enable PII detection and redaction: Bedrock can automatically detect and mask personal information including names, email addresses, phone numbers, Social Security numbers, and credit card numbers.
6. Configure Contextual grounding checks: these verify that the model's response is grounded in the provided context (from Knowledge Bases) and is relevant to the user's question. This reduces hallucination.
7. Optionally enable Automated Reasoning: this uses formal logic to achieve 99% accuracy on factual validation (AWS documentation). It is the strongest anti-hallucination control available.
8. Click Create, then attach the guardrail to your model invocations or agents.

Two tiers are available: Standard (newer, recommended) and Classic. The ApplyGuardrail API lets you use Bedrock Guardrails with models running outside Bedrock, so you can add safety filtering even if your inference runs on SageMaker, self-hosted models, or third-party APIs. For a full walkthrough of every policy type and configuration option, see our Bedrock Guardrails deep dive.

Step 7: Optimizing for Production

Once your prototype works, these optimizations reduce cost and latency before you go live. Robinhood achieved an 80% cost reduction after scaling from 500 million to 5 billion tokens per day on Bedrock in 6 months (AWS case study).

Prompt Caching

If your prompts share a common prefix (system instructions, few-shot examples), prompt caching with 5-minute and 1-hour TTL windows reduces costs by up to 90% and latency by 85% (AWS Bedrock Prompt Caching documentation). Enable it by adding the cacheConfig parameter to your Converse API calls.

Intelligent Prompt Routing

This feature automatically selects the most cost-effective model for each request based on complexity. AWS claims up to 30% cost reduction from routing alone (AWS Bedrock Intelligent Prompt Routing documentation). Enable it in the Bedrock console under inference settings.

Model Distillation

Compress a large model's capabilities into a smaller, cheaper one. AWS claims up to 500% faster inference and 75% lower cost with distilled models (AWS Bedrock Model Distillation documentation). Useful when you have validated that a smaller model can handle your specific use case with acceptable quality.

Batch Inference

For non-real-time workloads (document processing, bulk classification, data extraction), batch inference runs at 50% off on-demand pricing. Submit a batch job through the Bedrock console or API and retrieve results when complete.

Service Tiers

Bedrock offers four pricing tiers: Standard (pay-per-token, no commitment), Flex (50% off Standard for async-friendly workloads), Priority (75% premium for guaranteed high throughput and lowest latency), and Reserved (custom term commitment for dedicated capacity). Start on Standard, move to Flex for batch workloads, and evaluate Reserved only when your monthly spend justifies the commitment.

Troubleshooting and FAQ

Next Step

Build a retrieval-augmented generation workflow end to end. Upload your organization's internal documentation to a Knowledge Base, attach it to a Bedrock Agent with Guardrails enabled, and test it with real questions from your team. Measure response accuracy against your documents and compare costs across Nova Micro, Claude Haiku, and Llama models to find the price-performance sweet spot for your use case. For the full platform analysis, see our What Is Amazon Bedrock breakdown.

Amazon Web Services, AWS, Amazon Bedrock, Amazon Nova, and related marks are trademarks of Amazon.com, Inc. Anthropic and Claude are trademarks of Anthropic, PBC. Meta and Llama are trademarks of Meta Platforms, Inc. All other trademarks belong to their respective owners. This article is an independent editorial product of Tech Jacks Solutions and is not affiliated with, sponsored by, or endorsed by Amazon Web Services.