The Mock Exam

Treat this as the real thing. Read the scenario. Pick an answer. Then check the explanation. If you get 8/10, you are ready.

Question 1: The Hallucination Fix

Scenario: You are running a customer support bot on Vertex AI. It recently told a customer that your product "Cures cancer," which is false and dangerous. You checked the prompt, and it seems fine. Question: What is the most immediate and effective way to prevent this specific type of high-risk output?

A. Increase the Temperature.
B. Enable Grounding with Google Search.
C. Configure Safety Filters to block 'Medical Advice' or 'Dangerous Content'.
D. Retrain the model on Wikipedia.

Answer: C. While Grounding helps accuracy, immediate prevention of dangerous topics is best handled by Safety Filters or specialized guardrails that block specific categories of harm.

Question 2: The Latency Issue

Scenario: Your app uses a RAG pipeline. It takes 15 seconds to answer a user. Users are leaving. You analyzed the logs: The Retrieval step takes 0.5s, but the Generation step (Gemini Pro) takes 14.5s. Question: Which change preserves the most accuracy while reducing latency?

A. Switch to Gemini Flash.
B. Remove the RAG step.
C. Increase the Top-K retrieval count.
D. Switch to an on-premise GPU.

Answer: A. Gemini Flash is designed specifically for high-volume, low-latency tasks. It is much faster than Pro. Removing RAG (B) destroys accuracy.

Question 3: The Creative Block

Scenario: Your marketing team uses Gemini to write blog posts. They complain the output is "boring" and "repetitive." Question: Which parameter adjustment should you recommend?

A. Decrease Temperature to 0.
B. Increase Temperature to 0.9.
C. Increase Safety Filters.
D. Decrease Output Token Limit.

Answer: B. High Temperature increases randomness and creativity. Low temperature makes it deterministic (boring).

Question 4: The Legacy Data

Scenario: You want to build a search engine for your scanned PDF contracts from the 1990s. They are images, not text. Question: Which model capability is critical here?

A. Code Generation (Codey).
B. Text-to-Speech (Chirp).
C. Multimodal Vision (OCR/Gemini Pro Vision).
D. Reinforcement Learning.

Answer: C. You need OCR (Optical Character Recognition) or Multimodal Vision capabilities to "read" the text inside the images of the scanned PDFs.

Question 5: The Global Deployment

Scenario: You are a German bank. You want to use Vertex AI, but strict laws state customer data must strictly remain in Frankfurt. Question: How do you configure this?

A. Use the global endpoint.
B. You cannot use Cloud AI; you must build on-premise.
C. Specify location="europe-west3" (Frankfurt) when initializing the Vertex AI client.
D. Use a VPN.

Answer: C. Google Cloud allows Data Residency controls by specifying the region. europe-west3 is Frankfurt.

Question 6: The "Goldilocks" RAG

Scenario: You are building a RAG bot. When you retrieve 3 documents, the answer is missing details. When you retrieve 50 documents, the answer is slow and confused. Question: What is the architectural term for the "Confusion" caused by too much irrelevant context?

A. Hallucination.
B. The "Lost in the Middle" phenomenon (Context Saturation).
C. Overfitting.
D. Underfitting.

Answer: B. When you stuff too much context into a prompt, models notoriously struggle to find information hidden in the middle ("Lost in the Middle").

Question 7: Fine-Tuning Strategy

Scenario: You have 10,000 examples of English-to-French translations. Gemini Pro is already good at translation. Question: Is Fine-Tuning recommended here?

A. Yes, it is the only way to translate.
B. No, because Gemini is already a Foundation Model trained on translation; try Few-Shot prompting first.
C. Yes, because it saves money.
D. No, because 10,000 examples is too few.

Answer: B. Always Prompt First. Fine-tuning strictly for general translation is usually a waste of money because the base model is already excellent at it.

Question 8: Responsible AI

Scenario: You are building a face-detection system to unlock doors. Question: Which "Harm" must you rigorously test for to ensure fairness?

A. Latency Harm.
B. Allocative Harm (Fairness across skin tones/genders).
C. Toxic Speech.
D. Copyright Infringement.

Answer: B. Biometric systems have a history of performing poorly on certain skin tones. This leads to Allocative Harm (denying entry to valid employees based on race).

Question 9: Value Identification

Scenario: You can build Project A (Automated Coding Assistant for 500 devs) or Project B (AI recipe generator for the company cafeteria). Question: Why is Project A the better "Strategic Bet"?

A. It creates "Creation Value" at scale for high-cost employees.
B. It is easier to build.
C. Recipes are controversial.
D. Project B is illegal.

Answer: A. This is an ROI calculation. 500 devs saving 10% time is millions of dollars. The cafeteria saves pennies.

Question 10: Agent Orchestration

Scenario: An AI agent needs to check stock, then calculate tax, then place an order. Question: What mechanism allows the Agent to "wait" for the stock check before calculating tax?

A. Parallel Processing.
B. Chain of Thought (Reasoning Loop).
C. Temperature = 1.
D. Random Guessing.

Answer: B. The Chain of Thought (or ReAct loop) allows the agent to reason step-by-step: "Step 1 done. Now I have the data for Step 2."

What's Next?

Congratulations! You have finished the course. The final step is the Capstone Project. We will design a full Enterprise AI solution from scratch.

Knowledge Check

?Knowledge Check

What is the minimum score required to 'pass' this mock exam?

Practice Exam: 10 Tough Questions