Exam Guide & Traps

← Domain 4 · Cheatsheet →

---

Exam Traps

Trap 1: Model Selection

Know which model for which scenario:

If question mentions...	Think...
Multimodal (text + images + video)	Gemini
Longest context (1M+ tokens)	Gemini 1.5 Pro
Fast, low latency	Gemini Flash
Open-source	Llama, Mistral
Google's open-weight	Gemma
Third-party (Anthropic)	Claude

Trap 2: Customization Order

Common mistake: Jumping to fine-tuning when simpler methods work.

Correct order:

1. Prompt Design (zero-shot/few-shot) — try first
2. Grounding/RAG — need current data
3. Fine-Tuning — only if 100+ examples and above methods fail

Examples:

• ❌ "Fine-tune Gemini for current news"

• ✅ "Use grounding with Google Search"

• ❌ "Fine-tune for better formatting"

• ✅ "Use few-shot prompting with examples"

Trap 3: Vertex AI Component Confusion

Component	Purpose	When to Use
Model Garden	Browse/select models	Exploring available models
Vertex AI Studio	Test prompts, adjust parameters	Rapid prototyping
Vector Search	High-scale embeddings search	RAG implementation
AutoSxS	Compare model outputs	Evaluating which model is better

Trap 4: Grounding vs Fine-Tuning

Need	Solution
Current events	Grounding (Google Search)
Company data	Grounding (BigQuery, Document AI)
Domain-specific behavior	Fine-Tuning (100+ examples)
Custom format	Few-shot prompting

Trap 5: Data Privacy

Key facts (exam favorites):

• ✅ Google does not train foundation models on customer data
• ✅ Data stays in your GCP project
• ✅ Respects IAM permissions
• ✅ Enterprise compliance (SOC 2, ISO 27001)

---

Gen AI Leader Decision Matrix

Use this logic for scenario-based questions:

Business Need	Recommended Solution
"We need a prototype by tomorrow morning"	Google AI Studio
"We need to summarize private company PDFs"	RAG / Vertex AI Search
"We want to lower latency for a mobile app"	Gemini Nano / Model Distillation
"The model keeps making up facts"	Grounding with Google Search
"We need full MLOps and versioning"	Vertex AI Studio

Decision Quick Reference

"Which Vertex AI component?"

Explore models → Model Garden
Test prompts → Vertex AI Studio
Build RAG → Vector Search (Matching Engine)
Compare models → AutoSxS
Generate images → Imagen 2 (in Studio)
Speech tasks → Chirp

"Which customization method?"

Simple task, no examples → Zero-shot prompting
Custom format, 3-5 examples → Few-shot prompting
Need current data → Grounding (Google Search)
Need company data → Grounding (BigQuery/Document AI)
Have 100+ examples → Supervised Fine-Tuning (SFT)

"Which Gemini model?"

Complex reasoning, long docs → Gemini 1.5 Pro (1M+ context)
General tasks → Gemini Pro
Fast responses, low latency → Gemini Flash
Maximum capability → Gemini Ultra

"How to implement RAG?"

1. Generate embeddings (Vertex AI Embeddings API)
2. Store in Vector Search (Matching Engine)
3. Retrieve similar documents
4. Ground LLM with retrieved context

---

Exam Day Reminders

Think Like This

For Model Selection:

• Multimodal? → Gemini
• Longest context? → Gemini 1.5 Pro (1M+)
• Open-source? → Llama, Mistral
• Google's open? → Gemma

For Customization:

• Always try: Prompt → Grounding → Fine-Tuning
• Current data? → Grounding
• 100+ examples? → Fine-Tuning

For Vertex AI:

• Browse models? → Model Garden
• Test prompts? → Vertex AI Studio
• RAG? → Vector Search
• Compare? → AutoSxS

For Data Privacy:

• Google does NOT train on your data
• Data stays in your project
• Respects IAM

---

Common Question Patterns

"What should you do first?"

Always: Start with simplest approach

1. Try zero-shot prompting
2. Try few-shot prompting
3. Try grounding/RAG
4. Only then fine-tune

"How to reduce hallucinations?"

• ✅ Use grounding (Google Search, BigQuery)
• ✅ Implement safety filters
• ✅ Human review for critical tasks
• ✅ Provide factual context

"How to evaluate models?"

• ✅ Use AutoSxS (Automatic Side-by-Side)
• ✅ Define evaluation criteria
• ✅ Let judge model compare outputs
• ✅ Measure against business metrics

---

← Domain 4 · Cheatsheet →