Generative AI on AWS @amazon-bedrock :: One map of the AWS services that build, ground, protect and run generative AI applications > AWS groups its generative AI offerings by job: access to foundation models, agents that take actions, grounding answers in your data, ready-to-use assistants, safety and governance, and accelerated infrastructure. Most developers start with Amazon Bedrock for models and add the other pieces as the application grows. > Choose a managed service (Amazon Bedrock) when you want models through an API without managing servers, and choose Amazon SageMaker AI when you need to train, fine-tune or host your own models with control over the infrastructure. > Several older services are now closed to new customers (for example Amazon Bedrock Agents Classic, Amazon Q Business and Amazon Kendra), so check each service page before you design around it. + Generative AI on AWS + What is Amazon Bedrock? + What is Amazon SageMaker AI? Foundation models :: Where the models come from > Foundation models are large pre-trained models that you call through an API or host yourself. AWS offers them through Amazon Bedrock, through its own Amazon Nova family, and through SageMaker AI for models you manage. > The main choice is how much you want to manage: serverless access (Bedrock) or your own endpoints and training jobs (SageMaker AI). + Models at a glance in Amazon Bedrock Amazon Bedrock @amazon-bedrock :: Fully managed, secure access to over 100 foundation models from leading AI companies, including Amazon > Amazon Bedrock supports 100+ foundation models from providers such as Amazon, Anthropic, DeepSeek, Moonshot AI, MiniMax, OpenAI and xAI. You call them through the bedrock-runtime endpoint with the Converse, InvokeModel and other APIs. + Amazon Bedrock pricing Amazon Nova @amazon-nova :: Amazon's own foundation models for text, images, video, documents and speech, accessed through Amazon Bedrock > Amazon Nova 2 models such as Nova 2 Lite (text, images, video and documents in, text out) and Nova 2 Sonic (speech) support up to 1 million tokens of context. You access Nova models through an API using Amazon Bedrock. + Amazon Nova Amazon SageMaker AI @amazon-sagemaker-ai :: Build, train and deploy ML models, including foundation models, on managed infrastructure > SageMaker AI is a fully managed machine learning service with tools for the whole lifecycle: data preparation, training, tuning, deployment and monitoring. Choose it over Bedrock when you need custom training or direct control of endpoints. Agents :: Systems where a model plans and takes actions with tools > An agent lets a foundation model reason over a task, call tools and APIs, and use memory. AWS now recommends Amazon Bedrock AgentCore as the platform for building, deploying and operating agents. > Amazon Bedrock Agents (now called Classic) is closed to new customers, so new designs should use AgentCore or code-defined agents. + Amazon Bedrock Agents Classic maintenance mode Amazon Bedrock AgentCore @amazon-bedrock-agentcore :: Platform to build, deploy and operate agents securely at scale, with any framework and any model > AgentCore is modular: services such as Runtime, Memory, Gateway, Identity, Code Interpreter, Browser and Observability work together or independently. It works with open source frameworks such as CrewAI, LangGraph, LlamaIndex and Strands Agents. Amazon Bedrock Agents Classic @amazon-bedrock :: The original managed agents in Bedrock (action groups, knowledge bases); closed to new customers since July 30, 2026 > Existing customers can keep using Agents Classic, which is in maintenance mode. AWS recommends migrating to AgentCore, where the managed harness lets you declare the model, tools and instructions. Ground answers in your data :: Retrieval augmented generation (RAG) and vector storage > RAG retrieves relevant passages from your own data and adds them to the prompt so the model answers from current, private information. The pieces are a data source, an embedding model, a vector store and a retrieval step. > Amazon Bedrock Knowledge Bases manages that pipeline for you, and the vector store can be OpenSearch, Aurora PostgreSQL, S3 Vectors or others. + Amazon Bedrock Knowledge Bases Amazon Bedrock Knowledge Bases @amazon-bedrock :: Managed RAG, from ingestion to retrieval, with citations > Bedrock Knowledge Bases offers managed knowledge bases (Bedrock runs ingestion, indexing and retrieval) and customer-managed knowledge bases (you choose the vector store). Retrieve returns source chunks, and RetrieveAndGenerate returns a generated answer with citations. Amazon OpenSearch Service @amazon-opensearch-service :: Managed OpenSearch clusters, usable as a vector store for knowledge bases > OpenSearch Service deploys and operates OpenSearch clusters, and Bedrock Knowledge Bases supports OpenSearch Serverless and OpenSearch Managed Clusters as vector stores. Amazon S3 Vectors @amazon-simple-storage-service :: Vector buckets and indexes in Amazon S3, purpose-built and cost-optimized for vector storage > S3 Vectors stores vectors in vector indexes inside vector buckets, with subsecond latency for infrequent queries and as low as 100 milliseconds for frequent ones. Choose it for large, cost-sensitive vector datasets. Amazon Aurora PostgreSQL @amazon-aurora :: Aurora PostgreSQL with pgvector as a knowledge base vector store > The Aurora cluster must support pgvector 0.5.0 or higher to serve as a Bedrock knowledge base. Choose it when your team already runs PostgreSQL. Amazon Kendra @amazon-kendra :: Managed enterprise search with a GenAI index; closed to new customers > Amazon Kendra is no longer open to new customers, and AWS points to Amazon Bedrock Knowledge Bases for similar capabilities. Existing customers can still use Kendra GenAI indexes with Bedrock knowledge bases. Ready-to-use assistants :: Assistants you configure rather than build > These services give end users a working assistant without custom model integration. Check availability first, because some are changing or closed to new customers. Amazon Q Developer @amazon-q :: Generative AI assistant for software development and AWS operations > Q Developer IDE plugins reach end of support on April 30, 2027, and AWS points to Kiro for similar capabilities. The console chat experience is now called the AWS assistant. Amazon Q Business @amazon-q :: Assistant that answers from enterprise data; closed to new customers > Amazon Q Business is no longer open to new customers, and AWS points to AWS Quick for similar capabilities. AWS Quick @amazon-quick :: AI workspace for creating documents, dashboards and apps and for delegating research > AWS Quick turns a conversation into a document, spreadsheet, deck, image or dashboard, and connects to your tools and data. It can also build small apps from a description. Safety and governance :: Controls around models and data > Safety and governance cover what a model may be asked or say, who can call it, and what evidence you keep. Bedrock Guardrails handle content, IAM handles access, and CloudTrail records API calls. Amazon Bedrock Guardrails @amazon-bedrock :: Configurable safeguards for prompts and responses > Guardrails offer content filters, denied topics, word filters, sensitive information filters, contextual grounding checks and Automated Reasoning checks. The ApplyGuardrail API lets you use them independently of model invocation. Model evaluation @amazon-bedrock :: Compare models and knowledge bases with programmatic, human or judge-model evaluation > Bedrock evaluations cover models and knowledge bases, including RAG sources outside Bedrock. Judge-model evaluations use a second LLM to score responses and explain the score. AWS Identity and Access Management @aws-identity-and-access-management :: Who can invoke which models, agents and knowledge bases AWS CloudTrail @aws-cloudtrail :: Audit log of API calls, including Amazon Bedrock calls > CloudTrail captures Bedrock API calls from the console and from code. Model invocation logging is separate and records prompts and responses to CloudWatch Logs or Amazon S3. Accelerated infrastructure :: Hardware for training and running models yourself > When you host your own models you pick hardware: GPU-based accelerated computing instances, or AWS AI chips (AWS Trainium and AWS Inferentia) programmed through the AWS Neuron SDK. Amazon EC2 accelerated instances @amazon-ec2 :: Instances with hardware accelerators for graphics and floating point work > Accelerated computing instances use hardware accelerators or co-processors to run floating point calculations and graphics processing more efficiently than CPUs. AWS Neuron @aws-neuron :: Developer stack for deep learning and generative AI on AWS Trainium and AWS Inferentia > Neuron includes a compiler, runtime, training and inference libraries, and tools for monitoring and profiling. It works with PyTorch and JAX.

Generative AI on AWS mindmap