Beyond Classification

The exam now includes Generative AI. The key decision is: Prompt vs. Tune vs. Train.

---

1. The Hierarchy of GenAI

1. Prompt Engineering (Zero-Shot/Few-Shot):
   • Constraint: Using a frozen model (Gemini Pro).
   • Pros: Instant, cheap.
   • Cons: Limited context window, no new knowledge.
2. RAG (Retrieval Augmented Generation):
   • Constraint: Using a frozen model + Vector Search.
   • Pros: Access to your private documents. Grounded (Fact-based).
3. Parameter Efficient Fine-Tuning (PEFT/LoRA):
   • Constraint: Updating adapter weights (0.1% of model).
   • Pros: Can learn a new "Style" or strict format (JSON).
4. Full Fine-Tuning:
   • Constraint: Updating all weights.
   • Pros: Massive behavior change.
   • Cons: Extremely expensive (Requires A100s).

---

2. Vertex AI Model Garden

Do not download Llama 2 from HuggingFace manually if you can avoid it. Model Garden offers "Click-to-Deploy" for open models.

• Foundation Models: Gemini, PaLM, Imagen.
• Open Source: Llama, Mistral, Bert.

---

3. Vertex AI Agent Builder

Formerly "Gen App Builder" or "Discovery Engine".

• Search: Build a Google Search for your corporate PDFs.
• Conversation: Build a Chatbot that references those PDFs.

Exam Tip: If the requirement is "Build a customer support chatbot on our manuals in 1 day," use Agent Builder. Do not train a custom LLM.

---

Knowledge Check