AIF-C01 Cheatsheet — AI & Generative AI Fundamentals, AWS Service Map, RAG & Prompt Patterns
March 28, 20269 min read
High-signal AIF-C01 reference: AI/ML terminology, generative AI concepts (tokens, embeddings, RAG), AWS services (Bedrock, SageMaker and core AI services), prompt engineering patterns, evaluation rubrics, responsible AI, and security/governance essentials.
On this page
Keep this page open while drilling questions. AIF‑C01 rewards clean definitions, best-fit service selection, and risk-aware design (hallucinations, privacy, prompt injection, responsible use).
Quick facts (AIF-C01)
Item
Value
Questions
65 total
Scoring
50 scored + 15 unscored (unscored items are not identified)
Retrieval quality usually dominates answer quality
Last-minute traps
Picking fine-tuning when RAG + prompt engineering already solves grounding.
Sending raw PII/secrets to model prompts or logs.
Ignoring prompt injection in agent/tool scenarios.
Treating model confidence as truth without retrieval evidence.
Target candidate and scope boundaries
Target depth is practitioner-level: up to ~6 months of AI/ML exposure on AWS.
You are expected to use AI/ML services, not build custom AI/ML systems from scratch.
Usually out of scope: coding models, feature engineering, hyperparameter tuning, building ML pipelines/infrastructure, deep math/statistics, and designing governance frameworks from zero.
0) Core mental model: a GenAI app (with RAG)
flowchart LR
U[User] --> A[App / API]
A -->|Prompt + context| FM[Foundation Model]
A -->|Embed query| E[Embeddings]
E --> VS[(Vector Store)]
VS -->|Top-k chunks| A
A -->|Policy filters| G[Guardrails / Moderation]
A -->|Logs/metrics| O[Observability]
A -->|AuthN/AuthZ| IAM[IAM / Identity]
RAG in one sentence: retrieve relevant private content, then ask the model to answer using only that content (ideally with citations).
Subset of AI: models learn patterns from data to make predictions/decisions.
Deep learning
ML with neural networks (often needs more data/compute; strong for vision/language).
Supervised learning
Learn from labeled examples (classification/regression).
Unsupervised learning
Find structure without labels (clustering, dimensionality reduction).
Reinforcement learning
Learn actions via rewards/penalties (policies).
Feature / label
Input signal vs correct output.
Training vs inference
Fit the model vs use the model to predict/generate.
Overfitting
Great on training data, poor on new data (memorization).
Data leakage
Training sees information it shouldn’t (inflates metrics).
Drift
Data or reality changes → performance decays over time.
Metrics (common, conceptual)
Use case
Useful metrics
What to watch for
Classification
Precision/recall/F1, ROC-AUC
Class imbalance; false positives vs false negatives
Regression
MAE/MSE/RMSE
Outliers; error tolerance
Ranking/retrieval
Precision@k / Recall@k
“Did we retrieve the right things?”
ML lifecycle (high level)
flowchart LR
P[Define problem + metric] --> D[Collect/prepare data]
D --> T[Train + tune]
T --> E[Evaluate]
E --> DEP[Deploy]
DEP --> M[Monitor + feedback]
M --> D
Common best answer patterns:
If you can’t define a metric or get data, ML is usually the wrong first move.
Production ML needs monitoring (quality/latency/cost) and retraining plans.
2) Generative AI fundamentals (Domain 2)
Key GenAI terms (must know)
Term
Exam-friendly meaning
LLM
Language model that generates text from prompts.
Tokens
Model “chunks” of text; drives cost/limits.
Context window
Max tokens model can consider in one request.
Embeddings
Numeric vectors that capture semantic meaning for similarity search.
Vector store
Database/index optimized for similarity search over embeddings.
RAG
Retrieve relevant data and include it in the prompt to ground answers.
Temperature / top-p
Controls randomness vs determinism.
Hallucination
Output that sounds plausible but isn’t supported by facts.
Prompt injection
Untrusted text attempts to override instructions (“ignore previous”).
Prompting vs RAG vs fine-tuning (decision table)
Need
Best starting point
Why
Better instructions/format
Prompt engineering
Fast, cheap, reversible
Fresh/private knowledge
RAG
Grounds answers in your content without retraining
Consistent style/behavior
Fine-tuning
Teach patterns; reduces prompt complexity
A completely new capability
Usually not AIF-C01 scope
Consider specialist ML work
GenAI limitations to recognize
Factuality isn’t guaranteed → use grounding/citations and “unknown” responses.
Context is limited → don’t paste entire corpora; retrieve and summarize.
Outputs can be unsafe/biased → add guardrails, evaluation, and human review paths.
Costs scale with tokens → control prompt size, choose smaller models when acceptable, cache repeated work.
3) AWS service map (what to pick when)
Foundation models and ML platforms
You need…
Typical AWS answer
Managed foundation model access for GenAI apps
Amazon Bedrock
Build/train/tune/deploy custom ML models
Amazon SageMaker AI
Business/dev GenAI assistants
Amazon Q Business / Amazon Q Developer
AWS foundation model family for text/image/video generation
Amazon Nova
Bedrock features to recognize quickly
Need
Typical AWS answer
Policy filtering and safety controls for model outputs
4) RAG: design notes that show up in exam scenarios (Domain 3)
RAG architecture (end-to-end)
flowchart TB
subgraph Ingestion
S3[(Docs in S3)] --> C[Chunk + clean]
C --> EMB1[Create embeddings]
EMB1 --> VS[(Vector store)]
end
subgraph Answering
Q[User question] --> EMB2[Embed query]
EMB2 --> VS
VS --> K[Top-k chunks]
K --> P[Prompt template: instructions + context]
P --> FM[Foundation model]
FM --> A[Answer + citations]
end
High-yield design choices
Chunking: smaller chunks improve precision; larger chunks improve context. The exam often wants “tune chunking for relevance.”
Citations: if the requirement says “trust” or “audit,” add citations/source links.
Freshness: if content changes often, prefer RAG over fine-tuning.
Privacy: don’t send more data than needed; redact PII; restrict who can retrieve what (multi-tenant boundaries).
5) Prompt engineering patterns (Domain 3)
Techniques you should recognize
Technique
What it does
When to use
Clear instructions + constraints
Reduces ambiguity
Most questions
Few-shot examples
Improves formatting/edge cases
Structured outputs
Delimiters
Separates instructions vs data
Untrusted input scenarios
Output schema
Produces predictable JSON
App integrations
Grounding instructions
Reduces hallucinations
RAG and knowledge tasks
Refusal/escalation
Safer behavior
Policy/safety constraints
Prompt template (practical)
1Goal: Answer the user question using ONLY the provided context.
2Context:
3<<<
4{retrieved_chunks}
5>>>
6Rules:
7- If the answer is not in the context, say "Insufficient context".
8- Provide 2-3 bullet citations (source titles/ids).
9Output format (JSON):
10{"answer":"...", "citations":[{"source":"...","quote":"..."}]}
11User question: {question}
Anti-prompt-injection rule of thumb
Treat user-provided text as data, not instructions. If the model is allowed to call tools/actions, use allowlists and scoped permissions.
