aws-bedrock-ai

When to Use

Apply this skill any time an agent must reason about, design for, or implement against Amazon Bedrock, AWS AI managed services (SageMaker, Comprehend, Textract, Rekognition, Transcribe, Translate, Polly, Kendra, Amazon Q), or the licensing and data-privacy guarantees that surround foundation models on AWS. Use it whenever the prompt names a Bedrock primitive (foundation model, knowledge base, agent, guardrail, prompt, flow), an inference mode (on-demand, provisioned throughput, batch, cross-region), a fine-tuning surface, or a regulated AI deployment concern.

Trigger conditions:

• The agent must pick a foundation model for a stated workload (latency, context, multimodality, language, cost band).
• The agent must wire a Knowledge Base for Amazon Bedrock (RAG) and decide on the vector store, chunking strategy, and retrieval policy.
• The agent must design an Agents for Bedrock (or AgentCore) flow and define action-group OpenAPI / Lambda contracts.
• The agent must wire Guardrails — content filters, denied topics, sensitive-info filters, contextual-grounding checks — and bind them by ID at invoke time.
• The agent must choose between on-demand, provisioned throughput, batch inference, latency-optimised inference, or cross-region inference profiles, with cost and quota implications.
• The agent must apply licensing or AUP constraints (Anthropic, Meta Llama, Mistral, Cohere, Amazon Nova / Titan, AI21, Stability) before committing a model choice.
• The agent must satisfy a regulated deployment (EU/UK/CA data residency, KMS-CMK at rest, VPC-endpoint-only at transit, no model-training opt-in).

Procedure

1. Classify the AI workload. Pick one of: chat / RAG / agent / batch text / classification / extraction / image / speech / translation / fine-tuned. The category drives model family selection (references/foundation-models.md) and the inference mode (references/inference-modes.md).
2. Pick the foundation model and pin the version. Use the canonical Bedrock model identifier (<provider>.<family>-<size>-<variant>:<version>, e.g. anthropic.claude-sonnet-4-5-20251110-v1:0). Floating aliases without a version suffix are forbidden in production. Confirm GA in the target region(s) via kb:get_regional_availability and kb:search_documentation for "Bedrock model availability".
3. Confirm model access. Bedrock requires explicit, per-account, per-region model-access opt-in for many providers (Anthropic, Meta, Mistral, Cohere, AI21, Stability, some Amazon models). Document in the design that a console (or bedrock:PutModelAccessAgreement equivalent) acceptance step precedes any IaC deployment.
4. Apply licensing and AUP. Each provider carries its own EULA layered on top of the AWS Customer Agreement. Check references/licensing.md for the per-provider AUP highlights (Anthropic Acceptable Use Policy, Meta Llama Community License + AUP, Mistral commercial terms, Cohere usage policy, Amazon Nova / Titan custom terms, AI21 commercial terms, Stability commercial / community split). Surface any ambiguity as an Open Question in the design rather than picking silently.
5. Pick the inference mode and quotas. Apply references/inference-modes.md: on-demand for variable, low-volume; cross-region inference profile for capacity smoothing across regions; provisioned throughput for predictable high-RPM with reserved tokens; batch inference for asynchronous bulk jobs (50% pricing discount on supported families); latency-optimised inference for sub-second TTFT on supported models. Cost and quota implications are non-trivial and must land in the cost ROM.
6. Wire data-privacy and security. Bedrock does not use customer prompt or completion data to train base foundation models. Always-on guarantee, but the design must still enforce: KMS-CMK encryption for Knowledge Base data sources, S3 buckets, and OpenSearch Serverless collections; VPC endpoints (com.amazonaws.<region>.bedrock-runtime, bedrock, bedrock-agent-runtime, bedrock-agent) when the application is VPC-only; no log-group containing prompt bodies without a redaction strategy. Apply aws-waf-security-skill for the broader security posture.
7. Wire IAM least privilege. A Bedrock invocation principal needs only bedrock:InvokeModel (or InvokeModelWithResponseStream) plus the resource ARN of the specific model. Knowledge-base callers also need bedrock:Retrieve / RetrieveAndGenerate and bedrock-agent-runtime:* for agents. Guard against confused-deputy with aws:SourceArn / aws:SourceAccount on cross-service trust policies. Apply references/iam-patterns.md.
8. Wire RAG correctly. Knowledge Bases for Amazon Bedrock supports OpenSearch Serverless (default), Aurora PostgreSQL with pgvector, Pinecone, Redis Enterprise Cloud, and MongoDB Atlas. Pick per references/knowledge-bases.md: OSS for default, Aurora pgvector for VPC-only with existing RDS posture, third-party for existing footprint. Chunking strategy (default, fixed, semantic, hierarchical) and embedding model (amazon.titan-embed-text-v2:0, cohere.embed-english-v3, cohere.embed-multilingual-v3) are explicit decisions in the design.
9. Wire Guardrails by ID. A Guardrail is configured separately, versioned, and bound at invoke time via guardrailIdentifier + guardrailVersion. Content filters (hate, insults, sexual, violence, misconduct, prompt-attack), denied topics, sensitive-info filters (PII regex + denylist), word filters, and contextual-grounding checks compose. The component contract must name the bound guardrail.
10. Cost ROM via cost:. Bedrock prices are per 1k input tokens and per 1k output tokens; provisioned throughput is per model unit per hour (1 mu = stated tokens/min); batch inference is roughly 50% off on-demand for supported models; cross-region inference profiles add inter-region inference but no extra per-token premium. Knowledge Base storage costs apply on the underlying vector store.