Generative AI with Amazon Bedrock (Draft)
Amazon Bedrock is the fully managed service that provides secure access to foundational models from leading AI companies, enabling to build and scale generative AI applicatios.
Bedrock has two endpoint families with 5 API styles.
1. Two Endpoint Families
bedrock-mantle
AWS recommends to use this whenever possible. Use this when you want OpenAI-compatible or Anthropic-compatible APIs.
It supports,
- Responses API
- Chat Completions API
- Anthropic Messages API
along with this, mantle API supports stateful conversation management, tools support, projects and workspaces.
bedrock-runtime
Use this when you want more direct control with the AWS Bedrock APIs. Supported APIs are,
- InvokeModel API
- InvokeModelWithResponseStream
- InvokeModelWithBidirectionalSteam
- Converse API
- ConverseStream API
2. Five API Styles
Responses API
Use this when:
- You are migrating from OpenAI.
- You want a newer OpenAI-compatible API.
- You want stateful/agentic style interactions.
- You want easier tool-use patterns.
Chat Completions API
Use this when:
-
You already have code using
client.chat.completions.create. - You manually manage chat history.
- You want a lightweight chat API.
- You want easier migration from older OpenAI-style apps.
Anthropic Messages API
Use this when:
- You are using Claude-specific patterns.
- Your team already knows Anthropic SDK.
- You want Claude request/response structure.
- You are migrating from direct Anthropic API to Bedrock.
Converse API
Use this when:
- You want one consistent API across multiple Bedrock models.
- You may switch between Claude, Nova, Llama, Mistral, etc.
- You want a clean chat/message abstraction.
- You are building AWS-native GenAI applications.
InvokeModel API
Use this when:
- You need model-specific request format.
- You are calling embeddings, image generation, or non-chat models.
- You want low-level control.
- Converse does not expose a specific model feature you need.
- You are dealing with raw JSON payloads.
Quick Decision Table
| Requirement | Pick |
|---|---|
| AWS-native production chat | Converse on bedrock-runtime |
| Raw model-specific payload | Invoke on bedrock-runtime |
| Existing OpenAI app migration | Chat or Responses on bedrock-mantle |
| Stateful/agentic OpenAI-compatible app | Responses on bedrock-mantle |
| Existing Anthropic SDK / Claude-native style | Messages on bedrock-mantle |
| Embeddings / reranking / image generation | Usually Invoke on bedrock-runtime |
| Video generation / long-running media job | StartAsyncInvoke |
| Real-time voice conversation | InvokeModelWithBidirectionalStream |
