Generative AI and Amazon Bedrock Fundamentals

Shunku
#AWS#Amazon Bedrock#Generative AI#Foundation Models

Amazon Bedrock is a fully managed service that provides access to high-performing foundation models (FMs) from leading AI companies through a unified API. This article covers the fundamentals of generative AI and how Bedrock enables you to build AI-powered applications.

What is Generative AI?

Generative AI refers to AI systems that can create new content—text, images, code, audio, and video—based on patterns learned from training data.

flowchart LR
    A["Input Prompt"] --> B["Foundation Model"]
    B --> C["Generated Output"]

    subgraph Examples
        D["Text → Text"]
        E["Text → Image"]
        F["Text → Code"]
    end

    style B fill:#8b5cf6,color:#fff

Key Characteristics

Characteristic Description
Creative Generates novel content, not just retrieves
Contextual Understands and responds to natural language
Versatile Handles multiple tasks without task-specific training
Probabilistic Outputs may vary for the same input

What is Amazon Bedrock?

Amazon Bedrock provides a single API to access multiple foundation models, enabling you to build generative AI applications without managing infrastructure.

flowchart TB
    subgraph Bedrock["Amazon Bedrock"]
        API["Unified API"]

        subgraph Models["Foundation Models"]
            Claude["Anthropic Claude"]
            Titan["Amazon Titan"]
            Llama["Meta Llama"]
            Mistral["Mistral AI"]
            Stable["Stability AI"]
            Cohere["Cohere"]
        end
    end

    App["Your Application"] --> API
    API --> Models

    style Bedrock fill:#232f3e,color:#fff
    style API fill:#ff9900,color:#000

Key Benefits

Benefit Description
Fully Managed No infrastructure to provision or manage
Model Choice Access multiple FMs through one API
Security Data stays within your AWS account
Customization Fine-tune models with your data
Integration Native AWS service integrations

Foundation Models in Bedrock

Amazon Titan

Amazon's own family of foundation models.

Model Capability Use Case
Titan Text Text generation Summarization, Q&A, chat
Titan Embeddings Vector embeddings Semantic search, RAG
Titan Image Generator Image generation Creative content
Titan Multimodal Embeddings Text + image embeddings Multimodal search

Anthropic Claude

Known for safety and helpfulness.

Model Context Window Best For
Claude 3.5 Sonnet 200K tokens Balanced performance
Claude 3 Opus 200K tokens Complex reasoning
Claude 3 Haiku 200K tokens Fast, cost-effective

Meta Llama

Open-weight models with strong performance.

Model Parameters Characteristics
Llama 3.1 8B 8 billion Fast inference
Llama 3.1 70B 70 billion High quality
Llama 3.1 405B 405 billion State-of-the-art

Mistral AI

Efficient European models.

Model Use Case
Mistral 7B General purpose
Mixtral 8x7B Expert mixture
Mistral Large Complex tasks

Stability AI

Image generation specialists.

Model Capability
Stable Diffusion XL High-quality image generation
Stable Image Core Fast image generation
Stable Image Ultra Photorealistic images

Core Bedrock Capabilities

1. Model Inference

flowchart LR
    A["Prompt"] --> B["InvokeModel API"]
    B --> C["Foundation Model"]
    C --> D["Response"]

    style B fill:#ff9900,color:#000

2. Knowledge Bases (RAG)

Connect FMs to your private data sources.

flowchart TB
    A["Documents"] --> B["Knowledge Base"]
    B --> C["Vector Store"]
    D["Query"] --> E["Retrieve & Generate"]
    C --> E
    E --> F["Grounded Response"]

    style B fill:#3b82f6,color:#fff

3. Agents

Build autonomous AI agents that can take actions.

flowchart TB
    A["User Request"] --> B["Agent"]
    B --> C["Plan Actions"]
    C --> D["Execute Tools"]
    D --> E["Return Results"]

    style B fill:#22c55e,color:#fff

4. Guardrails

Implement responsible AI safeguards.

Guardrail Function
Content Filters Block harmful content
Denied Topics Restrict specific subjects
PII Filters Redact sensitive information
Word Filters Block specific words

5. Model Customization

Fine-tune models with your data.

Method Description
Continued Pre-training Adapt to domain-specific language
Fine-tuning Train on task-specific examples

Pricing Model

Bedrock offers flexible pricing options.

Option Description Best For
On-Demand Pay per token Variable workloads
Batch Discounted batch processing Large offline jobs
Provisioned Throughput Reserved capacity Consistent workloads

Getting Started

1. Enable Model Access

AWS Console → Amazon Bedrock → Model Access → Enable Models

2. First API Call (Python)

import boto3
import json

# Create Bedrock Runtime client
client = boto3.client('bedrock-runtime', region_name='us-east-1')

# Invoke Claude model
response = client.invoke_model(
    modelId='anthropic.claude-3-sonnet-20240229-v1:0',
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "max_tokens": 1024,
        "messages": [
            {"role": "user", "content": "Hello, Claude!"}
        ]
    })
)

# Parse response
result = json.loads(response['body'].read())
print(result['content'][0]['text'])

3. Using Converse API (Unified Interface)

response = client.converse(
    modelId='anthropic.claude-3-sonnet-20240229-v1:0',
    messages=[
        {"role": "user", "content": [{"text": "Hello!"}]}
    ],
    inferenceConfig={
        "maxTokens": 1024,
        "temperature": 0.7
    }
)

print(response['output']['message']['content'][0]['text'])

Model Selection Guide

flowchart TB
    A["What's your use case?"] --> B{"Text or Image?"}
    B -->|Text| C{"Complexity?"}
    B -->|Image| D["Stable Diffusion / Titan Image"]

    C -->|Complex reasoning| E["Claude Opus / Llama 405B"]
    C -->|General tasks| F["Claude Sonnet / Llama 70B"]
    C -->|Fast & cheap| G["Claude Haiku / Llama 8B"]

    style A fill:#3b82f6,color:#fff

Best Practices

Practice Recommendation
Start simple Begin with on-demand pricing
Test models Evaluate multiple models for your task
Use Converse API Unified interface across models
Enable logging Track usage and debug issues
Implement guardrails Add safety measures from the start

Key Takeaways

  1. Bedrock provides unified access to multiple foundation models via single API
  2. No infrastructure management - fully managed service
  3. Multiple capabilities - inference, RAG, agents, guardrails
  4. Data stays private - models don't train on your data
  5. Choose the right model - balance quality, speed, and cost

Next Steps

  • Set up model access in your AWS account
  • Try the Bedrock playground for interactive testing
  • Learn model-specific prompting techniques
  • Explore Knowledge Bases for RAG applications

References

Back to Blog