Amazon Bedrock
AWS Bedrock’s official service name is Amazon Bedrock. It is a fully managed Amazon Web Services service that gives access to foundation models (FMs) from multiple AI companies through a single API, so applications can experiment with, customize, and deploy generative AI without managing any model-hosting infrastructure.
AI Practitioner focus
- High-yield features are FM/model selection, token-based inference, Knowledge Bases for RAG, Guardrails, Prompt Management/versioning, Model Evaluation, model customization, and on-demand versus provisioned capacity.
- RAG/in-context learning do not change model weights; fine-tuning and continued pretraining do. Model distillation trains a smaller student to reduce cost/latency.
- For new tool-using agents, use Amazon Bedrock AgentCore; Bedrock Agents Classic is maintenance-only. Bedrock provides models/GenAI building blocks, while AgentCore provides production agent infrastructure.
Key points
- Offers foundation models from providers including AI21 Labs, Amazon (Titan, Nova), Anthropic, Cohere, DeepSeek, Luma AI, Meta, Mistral AI, OpenAI, Stability AI, TwelveLabs, and Writer, all callable through one consistent API rather than separate provider integrations.
- Knowledge Bases implement retrieval-augmented generation (RAG): they manage embeddings, vector storage, and re-ranking, connect to sources like S3, SharePoint, Confluence, Google Drive, and OneDrive, and expose a
RetrieveAndGenerateAPI that returns answers with citations. - Amazon Bedrock AgentCore is the current platform for building, connecting, and deploying AI agents (compute, memory, identity, and observability handled for you). The original Bedrock Agents feature has been renamed Bedrock Agents Classic and, as of July 30, 2026, is closed to new customers — AWS recommends AgentCore for any new agent development.
- Guardrails apply configurable safeguards (content filters, PII redaction, prompt-attack detection) across both prompts and model responses, independent of which underlying FM is used.
- Supports model evaluation (automatic scoring on accuracy, toxicity, and robustness, plus human evaluation workflows) to compare models before committing to one.
- Supports fine-tuning with labeled data and, for Amazon Titan models, continued pretraining on unlabeled domain data, as well as Provisioned Throughput for reserved, consistent-latency inference capacity instead of on-demand pricing.
- The boundary that matters: Bedrock provides managed access to pre-built foundation models for generative AI (text, chat, image, embeddings) through an API. Amazon Sage Maker is the platform for building, training, and deploying custom ML models from scratch, including your own infrastructure and algorithms — Bedrock is the generative-AI counterpart to SageMaker’s general-purpose ML tooling.