GCP-GAIL: Last-Minute Refresher
Final Review
This page is for the 15-minute "cram" session. Focus on the distinction between Grounding (accuracy) and Tuning (behavior).
🏗️ Domain 1: Vertex AI Foundation (~30%)
The Gemini Family
- Gemini Ultra: Largest, most capable model for highly complex reasoning and coding.
- Gemini Pro: Best-of-breadth; optimized for scaling across a wide range of text and video tasks.
- Gemini Flash: Optimized for speed and efficiency; ideal for high-volume, low-latency tasks.
- Gemini Nano: Designed for on-device efficiency (Android/Pixel).
Core Parameters
| Parameter | Effect | Use Case |
|---|---|---|
| Temperature | Controls randomness | High (0.8+) for creative writing; Low (0.1) for technical data. |
| Top-K | Limits vocabulary to K words | Prevents the model from picking highly unlikely "long tail" words. |
| Top-P | Dynamic vocabulary based on probability | Samples from the smallest set of words whose cumulative probability is P. |
🎯 Decision Trees
When to use RAG vs. Fine-Tuning?
Does the model need access to real-time or private data?
├─ Yes, and data changes daily → Use RAG (Vector Search + Grounding)
├─ Yes, but data is static and specialized → Use Fine-Tuning
└─ No, I just need a specific output format → Use Few-shot PromptingChoosing a Model Customization Path
What is the goal?
├─ Optimize for cost/latency? → Model Distillation
├─ Adopt a specific "voice" or persona? → Supervised Fine-Tuning (SFT)
└─ Prevent specific types of hallucinations? → Grounding with Google Search🛡️ Responsible AI & Safety
Safety Filters
Vertex AI provides adjustable thresholds for:
- Hate Speech
- Harassment
- Sexually Explicit
- Dangerous Content
Exam Tip: If a model refuses to answer a valid query, check if the Safety Filter thresholds are set too "Strict."
🔑 Key Acronyms
| Acronym | Full Form | Quick Definition |
|---|---|---|
| RAG | Retrieval-Augmented Generation | Attaching a "database" to the LLM to give it facts. |
| CoT | Chain of Thought | Telling the model to "Think step-by-step." |
| SFT | Supervised Fine-Tuning | Training a model on specific prompt-response pairs. |
| RLHF | Reinforcement Learning from Human Feedback | Training based on human "thumbs up/down" preferences. |
You've Got This!
Trust your knowledge of the Vertex AI workflow: Discover (Model Garden) → Experiment (Studio) → Customize (Tuning/RAG) → Deploy (Endpoints). Good luck! 🍀