Master AWS Bedrock vocabulary — model invocation patterns, guardrails, knowledge bases, and agentic orchestration.
0 / 5 completed
1 / 5
During standup, an AWS ML engineer explains the difference between InvokeModel and the Converse API. What's the key distinction?
InvokeModel requires you to format the request body per provider (Anthropic, Cohere, Llama formats all differ). The Converse API uses a single unified message format across all supported models — swap the model ID and the same code works. Both support streaming via their respective streaming variants.
2 / 5
In a PR review, a colleague adds Bedrock Guardrails to the invocation. What do guardrails provide?
Bedrock Guardrails are standalone resources (with their own ARN) that define content policies: topic denial, PII detection and redaction, hate speech filters, and custom word lists. You reference them by guardrailId in your API call and they apply to both the prompt and the model's response, consistently across any model.
3 / 5
An incident shows a Bedrock Knowledge Base returning irrelevant documents. How does the retrieval pipeline work?
Bedrock Knowledge Bases use a RAG pipeline: S3 source → chunking → embedding model (e.g. Titan Embeddings) → vector store (OpenSearch Serverless, Pinecone, Aurora pgvector, etc.) → semantic retrieval. Irrelevant results typically indicate a mismatch in chunking strategy (too large/small), wrong embedding model, or insufficient metadata filtering.
4 / 5
In a design review, the team evaluates Agents for Bedrock. What do they enable?
Agents for Bedrock implement the ReAct (Reason + Act) loop: the model reasons about the goal, selects an action group (Lambda + OpenAPI schema), executes it, observes the result, and continues until the goal is met. Session state persists across turns. You define capabilities via OpenAPI schemas — no hardcoded orchestration logic.
5 / 5
During a code review, a teammate uses Bedrock Prompt Management. What does it enable?
Bedrock Prompt Management stores prompt templates as versioned resources. Each version gets an ARN you reference in API calls. This separates prompt engineering from deployment cycles — you can update a prompt, version it, A/B test variants, and roll back without redeploying application code.