AIP-C01: Cheatsheet

← Overview · ← Exam Guide · Visual Study Kit →

Exam Day Reference

Review this page 5 minutes before the exam.

---

Foundation Model Quick Reference

FM	Vendor	Key Strength	Best For
Claude	Anthropic	200k token context, reasoning	Long docs, complex reasoning
Llama	Meta	Open-source, fine-tunable	Custom fine-tuning
Mistral	Mistral AI	Efficient, fast	Cost-efficient inference
Titan	AWS	AWS-native, embeddings	RAG embeddings, summarization
Cohere Embed	Cohere	Multilingual embeddings	Multilingual RAG

---

Bedrock API Comparison

API	Delivery	Use Case
InvokeModel	Synchronous (full response)	Simple query-response
InvokeModelWithResponseStream	Streaming (token by token)	Low-latency UX / chat
InvokeAgent	Streaming + trace	Multi-step agentic workflows

InvokeModel: Uses model-specific JSON in the request body. The expected schema varies by model family.

Converse API: Use when you want a consistent message-based interface across Bedrock models with optional inference settings, tool config, guardrails, and prompt variables.

Inference settings:
Low temperature = more deterministic and consistent
Higher temperature = more creative and variable
Lower topP = tighter token selection

---

Agents vs. Knowledge Bases

	Knowledge Base (RAG only)	Bedrock Agents
External API calls	No	Yes (via Action Groups + Lambda)
Multi-step reasoning	No	Yes
Document retrieval	Yes	Yes (optional)
Best for	Static knowledge Q&A	Dynamic workflows, tool use

---

High-Value Service Map

Service / Feature	Purpose	Quick Refresher
Amazon Bedrock	Serverless FM access	The central managed entry point for Claude, Titan, Llama, and other supported models
Knowledge Bases	Automated RAG	Handles ingestion, chunking, embedding, storage, and retrieval for private data
Bedrock Agents	Multi-step reasoning	Best when the workflow must use tools or take actions through Action Groups
Bedrock Flows	Orchestrated workflows	Better than Agents when the path is deterministic and explicitly defined
Prompt Caching	Cost + latency optimization	Reuses a repeated static prompt prefix such as a long system prompt
Amazon OpenSearch Serverless / Vector Engine	Semantic search	Stores embeddings and supports low-latency similarity search for RAG

---

Guardrails — Four Filter Types

Filter Type	What It Controls
Content Filters	Harmful categories: hate, violence, sexual content, insults
Denied Topics	Topics the model must refuse to discuss
Word Filters	Exact word/phrase blocklists
PII Redaction	Names, emails, SSNs, credit cards — Redact or Block

Key rules:

• Guardrails apply to both inputs AND outputs
• Must be explicitly applied per API call via guardrailIdentifier + guardrailVersion
• PII modes: Redact (mask with placeholder) vs. Block (reject request/response)
• Denied Topics = business-policy blocking at the topic level, not just exact keyword blocking

---

Safety, Governance, and Ops

Service / Feature	Purpose	Quick Refresher
Bedrock Guardrails	Content filtering	Filters PII, harmful content, and denied topics across model inputs and outputs
Model Evaluation	Quality control	Uses human or automated metrics such as Groundedness, Relevance, Accuracy, and Fluency
IAM Policies	Access control	Scope bedrock:InvokeModel permissions tightly; Bedrock does not train on your data by default
AWS PrivateLink	Data privacy	Keeps Bedrock traffic between your VPC and AWS backbone instead of the public internet
CloudWatch	Monitoring	Token counts, latency, alarms, dashboards
AWS X-Ray	Tracing	End-to-end tracing across multi-service GenAI request paths

---

Vector Store Options

Store	Type	Use When
Amazon OpenSearch Serverless	Managed, serverless	Bedrock Knowledge Bases (default)
Aurora PostgreSQL (pgvector)	RDS extension	Existing PostgreSQL infrastructure
Amazon Kendra	Enterprise search	NLP-powered enterprise retrieval

Retrieval concepts to remember:

• k-NN = nearest vectors
• Hybrid search = semantic + keyword
• BM25 = keyword ranking
• Reranking = re-score first-pass results
• Recall = find more relevant chunks
• Precision = reduce irrelevant chunks

---

Chunking Strategies

Strategy	Best For	Trade-off
Fixed-size	Uniform documents	May break context at boundaries
Fixed-size + overlap	Preserving cross-boundary context	Higher storage cost
Semantic	Varied, long-form content	Higher processing complexity
Hierarchical	Complex docs: broad + fine retrieval	More complex retrieval logic

---

Provisioned Throughput vs. On-Demand

	Provisioned Throughput (PTU)	On-Demand
Traffic	Predictable, 24/7	Sporadic, variable
Pricing	Fixed (per MU/hour)	Per token
Commitment	1 month or 6 months	None
Best for	Production steady-state	Dev/test, bursts

Batch Inference = ~50% cheaper than on-demand for non-real-time high-volume jobs.

---

Optimization Concepts

Concept	Purpose	Quick Refresher
Provisioned Throughput	Guaranteed capacity	Use for steady production workloads to avoid throttling
Model Distillation	Cost optimization	Train a smaller student model to mimic a larger teacher model
Temperature	Randomness control	0.0 = deterministic; higher values = more creative
Prompt Caching	Repeated-prefix reuse	Saves cost and latency when large static prompt prefixes repeat across requests

---

Model Evaluation Metrics

Metric	What It Measures
Groundedness	Response supported by retrieved context? (detects hallucinations)
Relevance	Response answers the user's question?
Accuracy	Factually correct?
Fluency	Well-written and natural?

---

Quick Decision Rules

Vector store for Knowledge Base?
→ Amazon OpenSearch Serverless (default) · pgvector on Aurora (alternative)

Multi-step reasoning + tool use?
→ Bedrock Agents + Action Groups (Lambda)

Need a unified message-based API across models?
→ Converse API

Content moderation / PII?
→ Guardrails for Amazon Bedrock

Lowest cost for repeated large prompt prefixes?
→ Prompt Caching

Highest grounding / accuracy for enterprise knowledge questions?
→ RAG with Knowledge Bases + strong retrieval

Audit trail for compliance?
→ AWS CloudTrail (not CloudWatch)

Operational monitoring (latency, errors, token counts)?
→ Amazon CloudWatch

Predictable 24/7 throughput?
→ Provisioned Throughput (PTU) — 1 or 6 month commitment

Bulk, non-real-time inference?
→ Batch Inference

Private connectivity to Bedrock API?
→ VPC Endpoint — bedrock-runtime for inference calls

Knowledge changes frequently / need traceability?
→ RAG (Knowledge Bases), not fine-tuning

Detect hallucinations in RAG?
→ Groundedness metric in Model Evaluation

Log all prompts and responses for AI governance?
→ Model Invocation Logging (to S3 or CloudWatch Logs)

Need a fixed multi-step workflow with retry/error handling?
→ Step Functions or Bedrock Flows, not Agents

---

Key Terminology

• FM: Foundation Model — pre-trained large AI model (Claude, Llama, Titan, etc.)
• RAG: Retrieval-Augmented Generation — FM inference + vector store retrieval
• PTU: Provisioned Throughput Unit — reserved Bedrock model capacity
• MU: Model Unit — unit of PTU capacity purchased
• PII: Personally Identifiable Information — data that identifies an individual
• Groundedness: Metric measuring how well a response is grounded in retrieved context
• Hallucination: FM generating information not present in the provided context
• Action Group: Lambda function exposed to a Bedrock Agent for tool use
• OpenSearch Serverless: Managed, serverless vector store used by Bedrock Knowledge Bases
• Batch Inference: Asynchronous bulk FM inference via S3 JSONL input/output
• Model Invocation Logging: Bedrock feature that logs all prompts + responses to S3/CloudWatch Logs
• Prompt Caching: Reuse of a static prompt prefix to reduce repeated compute and cost
• LLM-as-a-Judge: Using one model to evaluate the quality of another model's output

---

← Overview · ← Exam Guide · Visual Study Kit →