Amazon Bedrock – Generative AI Service

Amazon Bedrock Overview

• Amazon Bedrock is a fully managed service that provides access to high-performing foundation models (FMs) from leading AI companies through a single API.
• Bedrock enables building and scaling generative AI applications without managing infrastructure or training models from scratch.
• All data remains private — Bedrock does NOT use customer data to train or improve base models.
• Supports text generation, image generation, embeddings, chat, and multi-modal use cases.

Foundation Models

• Amazon — Nova (Micro, Lite, Pro, Premier), Titan (Text, Embeddings, Image, Multimodal)
• Anthropic — Claude (Haiku, Sonnet, Opus) family
• Meta — Llama 3.x and Llama 4 models
• Mistral AI — Mistral Large, Mistral Small
• Cohere — Command R, Command R+, Embed
• AI21 Labs — Jamba models
• Stability AI — Stable Diffusion (image generation)
• DeepSeek — DeepSeek-R1 (reasoning model)
• Models are accessed via the InvokeModel API — no need to provision instances.
• Cross-Region Inference — automatically routes requests to available regions for higher throughput.
• Inference Profiles — predefined configurations for consistent model behavior.

Amazon Bedrock Agents

• Build autonomous AI agents that can plan, orchestrate, and execute multi-step tasks.
• Agents can invoke APIs, query databases, and interact with enterprise systems.
• Action Groups — define what actions an agent can take (Lambda functions, API schemas).
• Knowledge Bases — give agents access to company data for RAG (Retrieval-Augmented Generation).
• Multi-agent collaboration — agents can delegate tasks to other specialized agents.
• Return of control — pause agent execution and return control to the application for human-in-the-loop workflows.
• Code Interpreter — agents can generate and execute code to perform calculations and data analysis.
• Memory — agents retain context across conversations for personalized interactions.

Amazon Bedrock Knowledge Bases

• Implements Retrieval-Augmented Generation (RAG) — connects FMs to company data sources.
• Data sources: S3, Confluence, SharePoint, Salesforce, Web Crawler, custom connectors.
• Vector stores: OpenSearch Serverless, Aurora PostgreSQL, Pinecone, Redis Enterprise, MongoDB Atlas, Neptune Analytics.
• Chunking strategies: Fixed-size, semantic, hierarchical, no chunking.
• Parsing: Built-in parsers for PDF, Word, HTML, Markdown, CSV, Excel.
• Advanced RAG:
  • Metadata filtering — filter results by document attributes
  • Hybrid search — combines semantic + keyword search
  • Re-ranking — uses a re-ranker model to improve result relevance
  • Query decomposition — breaks complex queries into sub-queries
• GraphRAG — uses knowledge graphs (Neptune) for relationship-aware retrieval.

Amazon Bedrock Guardrails

• Implement safeguards for generative AI applications — works with any FM on Bedrock or custom models.
• Content filters — block harmful content categories (hate, insults, sexual, violence, misconduct) with configurable thresholds.
• Denied topics — define topics the model should refuse to discuss.
• Word filters — block specific words, phrases, or profanity.
• Sensitive information filters (PII) — detect and redact/mask PII (names, SSN, credit cards, etc.).
• Contextual grounding checks — detect hallucinations by verifying responses against source material.
• Automated Reasoning checks — uses formal logic to validate factual accuracy.
• Guardrails can be applied to both inputs and outputs.
• Works with Bedrock Agents, Knowledge Bases, and direct InvokeModel calls.
• ApplyGuardrail API — apply guardrails to any text, even outside Bedrock.

Amazon Bedrock Model Customization

• Fine-tuning — train a model on your specific data to improve performance for your use case.
• Continued Pre-training — train a model on domain-specific unlabeled data for deeper domain knowledge.
• Model Distillation — transfer capabilities from a larger teacher model to a smaller, faster student model.
• Custom models are private — only accessible in your account.
• Training data stored in S3, encrypted with KMS.
• Provisioned Throughput — purchase dedicated capacity for custom or base models for consistent performance.

Amazon Bedrock Model Evaluation

• Compare model performance using automatic evaluation (built-in metrics) or human evaluation (human reviewers).
• Automatic metrics: accuracy, robustness, toxicity, BERTScore, ROUGE.
• LLM-as-a-judge — use a foundation model to evaluate outputs of other models.
• Compare multiple models side-by-side for your specific use case.
• Results stored in S3 for analysis.

Amazon Bedrock Flows

• Visual workflow builder for creating generative AI pipelines.
• Chain prompts, knowledge bases, agents, guardrails, and Lambda functions into workflows.
• Supports conditional branching, parallel execution, and iterative loops.
• Version and deploy flows independently.

Amazon Bedrock Prompt Management

• Create, version, and manage prompts centrally.
• Prompt variables — use placeholders for dynamic content.
• Prompt Caching — cache context for frequently used long prompts to reduce latency and cost.
• Intelligent Prompt Routing — automatically routes requests to the optimal model based on prompt complexity.

Amazon Bedrock Studio

• Web-based playground for non-technical users to build and test generative AI applications.
• Create projects with shared resources (agents, knowledge bases, guardrails).
• SSO integration via IAM Identity Center for team collaboration.

Amazon Nova Models

• Amazon’s own family of foundation models, purpose-built for Bedrock.
• Nova Micro — text-only, lowest latency, lowest cost (ideal for simple tasks).
• Nova Lite — multimodal (text, image, video input), fast and cost-effective.
• Nova Pro — multimodal, best balance of accuracy, speed, and cost.
• Nova Premier — most capable, best for complex reasoning and agentic workflows.
• Nova Canvas — image generation with watermark detection.
• Nova Reel — video generation (up to 6 seconds).
• Nova Sonic — speech-to-speech model for natural conversations.

Bedrock Security

• Data privacy — customer data is NOT used to train base models; model inputs/outputs are not shared.
• Encryption — data encrypted in transit (TLS 1.2+) and at rest (KMS). Customer-managed keys supported.
• VPC connectivity — access Bedrock via VPC endpoints (PrivateLink) for private network traffic.
• IAM integration — fine-grained access control with IAM policies, resource-based policies.
• Model access control — explicitly enable which models are available in your account.
• CloudTrail logging — all API calls logged for auditing.
• Model Invocation Logging — log prompts and responses to S3/CloudWatch for compliance.
• Service Control Policies — restrict Bedrock usage at the organization level.

Bedrock Pricing

• On-Demand — pay per input/output token (no commitment, most flexible).
• Batch Inference — up to 50% cheaper for non-time-sensitive workloads.
• Provisioned Throughput — reserved capacity with committed model units (1-month or 6-month terms).
• Model Customization — charged per training token processed.
• Knowledge Bases — charged per storage and retrieval query.
• Guardrails — charged per 1,000 text units processed.

AWS Certification Exam Practice Questions

1. A company wants to build a customer support chatbot that can access company documentation stored in S3 to answer questions accurately. Which Bedrock feature should they use?
   1. Fine-tuning
   2. Knowledge Bases (RAG)
   3. Guardrails
   4. Model Evaluation
2. An organization needs to ensure their generative AI application never discusses competitor products and redacts any PII from responses. Which Bedrock feature provides this?
   1. Knowledge Bases
   2. Model Customization
   3. Guardrails (denied topics + PII filters)
   4. Prompt Management
3. A startup wants to reduce the latency and cost of their Bedrock application that uses Claude for simple classification tasks. Which approach is most cost-effective?
   1. Provisioned Throughput for Claude
   2. Fine-tune Claude on classification data
   3. Use Intelligent Prompt Routing or switch to Nova Micro
   4. Use Batch Inference
4. A financial services company requires that all Bedrock API traffic stays within their private network and all prompts/responses are logged for regulatory compliance. Which features should they enable?
   1. CloudTrail + S3 encryption
   2. VPC endpoints (PrivateLink) + Model Invocation Logging
   3. Guardrails + Knowledge Bases
   4. IAM policies + Batch Inference
5. A company wants an AI agent that can look up customer orders in DynamoDB, check shipping status via an API, and send email notifications. Which Bedrock feature enables this?
   1. Knowledge Bases
   2. Bedrock Flows
   3. Bedrock Agents with Action Groups
   4. Model Customization
6. An enterprise needs to transfer the capabilities of a large, expensive model to a smaller model for production use to reduce inference costs while maintaining quality. Which feature supports this?
   1. Fine-tuning
   2. Continued Pre-training
   3. Model Distillation
   4. Provisioned Throughput

Related Posts

• AWS AI & ML Services Cheat Sheet
• AWS SageMaker
• AWS Lambda vs Fargate vs App Runner
• AWS Serverless Services Cheat Sheet

References

Amazon Bedrock User Guide

Amazon Bedrock Agents

Amazon Bedrock Knowledge Bases

Amazon Bedrock Guardrails

Amazon Bedrock Pricing