Amazon Bedrock mindmap

Amazon Bedrock

Fully managed service that offers foundation models from Amazon and leading AI companies through APIs

• Models: Which models you can call and how you get access
  • Model catalog: Model cards for Amazon, Anthropic, Meta, Mistral AI, OpenAI, xAI and many other providers
  • Pay per token, serverless: On-demand inference with no infrastructure to manage
  • Amazon Bedrock Marketplace: Over 100 more specialized models, deployed to endpoints managed by SageMaker AI
  • Custom model import: Import a customized open-source model, or a SageMaker AI fine-tuned Amazon Nova model
• Inference: How you send requests and trade off cost, latency and capacity
  • Converse API: One unified interface across models that support messages
  • InvokeModel API: Direct, model-specific request and response format
  • OpenAI-compatible APIs: Responses and Chat Completions APIs on the bedrock-runtime endpoint
  • Service tiers: Reserved, Priority, Standard and Flex
  • Cross-Region inference: Inference profiles route requests across Regions for higher throughput
  • Batch inference: Asynchronous processing of many prompts from and to Amazon S3
  • Provisioned Throughput: Model units at a fixed hourly price, with optional commitment
  • Prompt caching: Reuse a cached prompt prefix to cut latency and input token cost
  • Intelligent prompt routing: Routes each request between models in the same family
• Customization: Adapt a model with your data
  • Supervised fine-tuning: Train on labeled examples to improve a task
  • Reinforcement fine-tuning: Improve a model with reward functions instead of labeled pairs
  • Model distillation: Transfer knowledge from a larger teacher model to a smaller student model
• Knowledge Bases: Managed retrieval augmented generation (RAG)
  • Data sources: Amazon S3, Confluence, Google Drive, OneDrive, SharePoint, web crawler or custom
  • Chunking: Default (about 300 tokens), fixed-size, hierarchical, semantic or none
  • Vector stores: OpenSearch Serverless and Managed Clusters, Neptune, Aurora, S3 Vectors, Pinecone, Redis Enterprise Cloud, MongoDB Atlas
  • Retrieve and RetrieveAndGenerate: Get chunks only, or a generated answer with citations
  • Structured data: Natural language to SQL over Amazon Redshift
  • Graph-based retrieval: Knowledge bases built on Amazon Neptune Analytics graphs
  • Reranking: Use reranking models to influence which retrieved results come back
• Guardrails: Configurable safeguards for inputs and outputs
  • Content filters: Hate, Insults, Sexual, Violence, Misconduct and Prompt Attack, each with a strength
  • Denied topics: Topics your application must avoid, defined in natural language
  • Word filters: Custom words and phrases plus a managed profanity list
  • Sensitive information filters: Block or mask PII types and custom regex patterns
  • Contextual grounding check: Flags responses not grounded in the source or irrelevant to the query
  • Automated Reasoning checks: Validate responses against a set of logical rules
  • ApplyGuardrail API: Evaluate content with a guardrail without invoking a model
• Build tools: Higher-level building blocks
  • Agents Classic: Action groups, knowledge bases, memory, code interpretation and multi-agent collaboration; closed to new customers
  • Flows: Visual workflows that link prompts, models, knowledge bases and AWS services
  • Prompt management: Create, version and reuse prompts with variables
  • Evaluations: Programmatic, human and judge-model evaluation of models and knowledge bases
  • Data Automation: Turn documents, images, video and audio into structured data
• Security and monitoring: Protect data and see what happens
  • Your data stays yours: Inputs and outputs are not used to train models or shared with model providers
  • Encryption: TLS in transit; AWS KMS keys for resources such as custom models, agents and ingestion jobs
  • Private connectivity: Interface VPC endpoints through AWS PrivateLink
  • Access control: IAM policies and roles for Bedrock resources
  • Invocation logging: Log requests and responses to CloudWatch Logs or Amazon S3
  • Audit: Bedrock API calls recorded in CloudTrail