AWS Bedrock & AI Services Study Guide — AI Intermediate

AWS Bedrock & AI Services

This domain accounts for 22% of the ai-intermediate exam — the single largest domain. It tests your ability to architect, configure, and troubleshoot AI applications built on Amazon Bedrock and AWS's managed AI services. Expect scenario-based questions requiring you to choose between foundation models, configure Knowledge Bases or Agents, apply Guardrails, select the right AWS AI service for a task, and reason about pricing and security trade-offs. Memorizing exact API names, component relationships, and service boundaries is essential.

Amazon Bedrock Overview: Foundation Models

Amazon Bedrock is a fully managed service that provides access to foundation models (FMs) from multiple providers via a single API, without requiring you to manage infrastructure or train models from scratch.

Model Providers and Use Cases

Key Concepts

• Model access must be explicitly requested per AWS account; not all models are available in all regions.
• Cross-region inference allows routing requests to models in other regions when the home region is at capacity (covered separately below).
• Bedrock is serverless — no EC2 instances, no model weights to manage.
• Titan Embeddings is the primary model for generating vector embeddings within Knowledge Bases.

Common traps:

• Confusing "requesting model access" with "deploying" a model — there is no deployment step.
• Assuming all Claude versions are always available in every region — regional availability varies.
• Selecting SageMaker when the question asks for managed FM access without custom training.

Bedrock API: InvokeModel, Converse API, Streaming, Cross-Region Inference

Core API Methods

InvokeModel vs Converse — Critical Distinction

• InvokeModel requires you to format the request body differently for each model family (e.g., Claude uses anthropic_version, Titan uses inputText). You must handle schema differences in your application.
• Converse abstracts these differences. The same code works across Claude, Titan, Llama, and Mistral. Prefer Converse for new multi-model applications.

Streaming

• Streaming is invoked via InvokeModelWithResponseStream or ConverseStream.
• The response is a stream of chunk events, not a single JSON blob.
• Use streaming when: end-user is waiting in real time, responses are long, or latency matters.

Cross-Region Inference

• Cross-region inference profiles allow Bedrock to automatically route requests to the same FM hosted in a secondary region when the primary region has insufficient capacity.
• You reference a system-defined inference profile (e.g., us.anthropic.claude-3-sonnet-20240229-v1:0) instead of a model ID.
• The prefix us. or eu. denotes the inference profile geography.
• Data residency — data may leave your primary region; evaluate compliance requirements before enabling.

Common traps:

• Using InvokeModel with a hardcoded Claude request body and then claiming the code is model-agnostic — it is not.
• Believing cross-region inference requires you to configure separate API endpoints — it does not; the profile ID handles routing transparently.
• Confusing ConverseStream (chat streaming) with InvokeModelWithResponseStream (raw model streaming) — both stream, but Converse is model-agnostic.