Module 16 cheat sheet
One-page revision summary for Generative AI on AWS for Banking: services, Azure mapping, CLI commands, limits, decision rules, exam traps and cost traps.
Assumes: AWS CLI v2 · AWS SDK for .NET v4 (AWSSDK.BedrockRuntime) · Region us-east-1 · facts as stated in the Module 16 lessons (checked 2026-11-18). Verified against official sources on 2026-11-18.
A printable summary of Module 16. Each row links back to the lesson with the detail. Model IDs, prices and availability change often: confirm them in the Bedrock console before you rely on them.
Key services
| Service | What it is for |
|---|---|
| Amazon Bedrock (Converse API) | Call models from many providers through one request shape; stateless, so history costs tokens each turn |
Inference profiles (us., eu., global.) | Route a request across Regions; Claude 5.5 models have no in-Region support in us-east-1 |
| Bedrock Knowledge Bases | Managed RAG: ingest approved documents, retrieve with citations; managed or customer-managed |
| Bedrock Guardrails | Content, prompt-attack, denied-topic, PII and grounding checks; ApplyGuardrail works with no model |
| Amazon Bedrock AgentCore | Hosting and governance for agents (Runtime, Gateway, identity, memory, policy, observability) |
| Kiro | Successor to Amazon Q Developer IDE plugins; Q in the console continues |
| Bedrock model invocation logging | A second copy of your prompts: classify, encrypt, set retention |
Bedrock Evaluations / fmeval | Evaluate foundation models, a model-risk validation input |
.NET / Azure → AWS
| Microsoft / Azure | AWS |
|---|---|
| Azure OpenAI / AI Foundry | Amazon Bedrock |
Azure.AI.OpenAI | AWSSDK.BedrockRuntime (Converse) |
Microsoft.Extensions.AI IChatClient | Same interface over Bedrock or a local model |
| Azure AI Search (RAG) | Bedrock Knowledge Bases |
| Content Safety, Prompt Shields | Bedrock Guardrails (prompt attack filter) |
| Azure AI Agent Service | Bedrock AgentCore |
| GitHub Copilot | Kiro |
Most useful CLI commands
| Task | Command |
|---|---|
| List models | aws bedrock list-foundation-models --profile learn-sandbox |
| Read a profile's destination Regions | aws bedrock get-inference-profile --inference-profile-identifier <id> |
| One-off model call | aws bedrock-runtime converse --model-id <profile-id> --messages file://msg.json |
| Create / version a guardrail | aws bedrock create-guardrail … · aws bedrock create-guardrail-version --guardrail-identifier <id> |
| Test text against a guardrail | aws bedrock-runtime apply-guardrail … |
| Delete a guardrail (teardown) | aws bedrock delete-guardrail --guardrail-identifier <id> |
| Sync a knowledge base | aws bedrock-agent start-ingestion-job … |
| Query a knowledge base | aws bedrock-agent-runtime retrieve … |
| Check logging config | aws bedrock get-model-invocation-logging-configuration |
| Anything left tagged? | aws resourcegroupstaggingapi get-resources --tag-filters Key=Project,Values=cea-lab |
Key limits and defaults
| Item | Value | Lesson |
|---|---|---|
| Claude 5.5 context / max output | 1M tokens / 128K | Bedrock basics |
| Titan Text Embeddings V2 | InvokeModel only; 8K-token input; 1,024, 512 or 256 dimensions | Bedrock basics |
| Prompt caching (5.5 cards) | 512-token minimum, up to 4 checkpoints, 5-minute or 1-hour life | Bedrock basics |
| Guardrail text unit | Up to 1,000 characters; 5,600 characters bills as 6 units | Guardrails |
| Guardrail prices per 1,000 units | Content filters and denied topics 0.15 USD; PII and grounding 0.10 USD; word and regex filters free | Guardrails |
| Denied-topic definition | 200 characters (Classic), 1,000 (Standard) | Guardrails |
| Managed KB embedding | 1,024 dimensions, float32; type cannot change after creation | Knowledge Bases |
Retention mode aws_review | Up to 30 days, may be reviewed by AWS; some models require it | Data privacy |
Decision rules
| If… | Then… |
|---|---|
Calling a Claude 5.5 model from us-east-1 | Use a us. or global. profile ID, not the bare model ID |
| Personal data and residency matter | Prefer a geographic profile or single Region; read destinations with GetInferenceProfile |
| Choosing a RAG store | Managed Knowledge Base first; customer-managed only for a specific store need |
| Users see different documents | Metadata filters built server-side from verified identity |
| Guardrail on a RAG call | Mark only the question for evaluation; scanning all passages can cost more than the model |
| Agent can change state | Allow-list tools, step and token limits, human approval for writes |
| Output affects a customer or money | Human review, maker-checker, default is not released |
| Need to block a model | IAM or SCP deny on bedrock:InvokeModel; access is automatic on first use |
Common exam traps
- Retired Model access page: commercial Regions enable models on first use; the control is IAM and SCPs.
- Closed services still appear in exam guides: Kendra, Q Business, Clarify, Model Monitor, Ground Truth, A2I (closed 2026-07-30). Replacements: Bedrock Knowledge Bases, Amazon Quick, SHAP and MLflow, Evidently, a self-built review queue.
- Bedrock Agents Classic is closed to new customers; use AgentCore. Models, Knowledge Bases and Guardrails are unaffected.
- Guardrails are probabilistic and text-only: tool results, tool definitions and tool-call arguments are not evaluated; the trace holds the unmasked match.
- Relevance is not permission: similarity search does not enforce access control.
- RAG is not fine-tuning: RAG adds documents at question time and needs no training.
- Prompt injection (OWASP LLM01): indirect injection through retrieved content is the main RAG risk.
Cost traps
| Trap | Avoid it by |
|---|---|
| Knowledge base left running | Paid while it exists: delete it the same day, or use local pgvector |
| Guardrail evaluating the whole prompt | Evaluate the question only; billing follows text units |
| Long prompts and chat history | History is resent each turn; long-context rates apply above a threshold |
| Provisioned Throughput | Billed hourly even when idle |
| Unbounded agent loops | Cap steps, tokens and time; cap MaxTokens on every call |
| Invocation logs | Retention and storage grow; set lifecycle rules and encrypt |
Independent learning resource — not affiliated with, endorsed or sponsored by Amazon Web Services. AWS and related marks are trademarks of Amazon.com, Inc. or its affiliates. All banking data in examples is fictional. Sample architectures are for learning and are not automatically compliant with GDPR, PCI DSS or any banking regulation.
Human-in-the-loop and use-case triage for GenAI
Classify GenAI use cases as good fits, controlled fits or do-not-proceed, choose the right human-review pattern, and build a review queue with maker-checker rules and a Step Functions approval flow in C#, CDK and Terraform.
Module 17: Daily-Work Playbooks & Career
Step-by-step diagnosis guides for the problems you will actually face in a bank running .NET on AWS, from duplicate payments to AccessDenied errors and surprise bills. The module closes with interview preparation, whiteboard practice and a guide to presenting your capstone projects as a portfolio.