Generative AI Leader Practice Exam 2 — 20 Questions

Question 1

What are THREE benefits of using embedding models in a Retrieval Augmented Generation (RAG) system?

Accepted Answer

They compress text into dense vectors for efficient retrieval.. Embedding models convert text into dense vector representations that capture semantic meaning, enabling efficient similarity search in vector databases. This compression reduces the dimensionality of the data, allowing the RAG system to quickly retrieve the most relevant documents from a large corpus based on vector distance metrics like cosine similarity.

Answer

They allow the model to generate new training data automatically.

Answer

They provide deterministic outputs for the same query.

Question 2

A company has deployed a GenAI-powered report generation system using Vertex AI. They notice that the cost is higher than expected. Investigation shows that many requests include very long prompts with repetitive boilerplate text. Which cost optimization strategy is MOST effective?

Accepted Answer

Enable context caching for repeated prompt prefixes. Caching repeated prompt prefixes can significantly reduce token usage and cost. Batch requests help with throughput but not with per-request token savings. Reducing model size may hurt quality. Ignoring is not a strategy.

Answer

Increase the batch size for batch requests

Answer

Switch to a smaller model size

Answer

Ignore the cost increase as it will stabilize

Question 3

Which TWO statements are true about generative AI models?

Accepted Answer

They are typically pre-trained on large datasets.. Generative AI models, such as GPT-4 or DALL-E, are typically pre-trained on vast, diverse datasets (e.g., terabytes of text or images) using unsupervised or self-supervised learning. This pre-training phase allows the model to learn statistical patterns, grammar, and world knowledge, which is then fine-tuned for specific tasks. Without this large-scale pre-training, the model would lack the foundational understanding needed to generate coherent and contextually relevant outputs.

Answer

They are deterministic by design.

Answer

They always produce the same output for the same input.

Answer

They require no data for training.

Question 4

A company is deploying a Gemini 1.0 Ultra model for a code generation assistant. They have set up Vertex AI Model Evaluation with a custom evaluation dataset to measure pass@1 accuracy. The initial evaluation shows 65% pass@1. They want to improve to 80% without collecting more training data. They have already attempted basic prompt engineering (e.g., 'write correct code') with limited improvement. Which approach is most likely to achieve the desired improvement?

Accepted Answer

Use chain-of-thought prompting with few-shot examples of correct code generation.. Chain-of-thought prompting with few-shot examples is the most effective approach because it guides the model through step-by-step reasoning, which is critical for complex code generation tasks. This technique leverages the model's in-context learning ability to improve accuracy without additional training data, directly addressing the need to boost pass@1 from 65% to 80%.

Answer

Reduce the temperature to 0 and set top_p to 1.

Answer

Increase the number of output tokens and enable beam search with width 4.

Answer

Apply reinforcement learning from human feedback (RLHF) using a reward model trained on the existing evaluation dataset.

Question 5

A team is using a pre-trained language model to summarize legal documents. They find that summaries often miss key dates and parties involved. Which technique would most effectively improve factual accuracy?

Accepted Answer

Fine-tune the model on a dataset of legal summaries with annotated key entities.. Fine-tuning on a dataset of legal summaries with annotated key entities directly teaches the model to recognize and reproduce critical factual elements like dates and parties. This supervised learning approach adjusts the model's weights to prioritize entity extraction and accurate generation, which is the most effective method for improving factual accuracy in domain-specific tasks.

Answer

Use top-p sampling with a low p value.

Answer

Increase the temperature parameter.

Answer

Use chain-of-thought prompting.

Question 6

A company has a large dataset of proprietary documents and wants to build a Q&A system using a foundation model without exposing the documents to the model. Which approach is most appropriate?

Accepted Answer

Use RAG with Vertex AI Vector Search and embeddings. Retrieval-Augmented Generation (RAG) with Vertex AI Vector Search allows the system to retrieve relevant document chunks via embeddings without exposing the full documents to the foundation model. The model only sees the retrieved context in the prompt, ensuring proprietary data remains isolated and not used for training or memorization.

Answer

Use a zero-shot model with context in prompt

Answer

Fine-tune the model on the documents

Answer

Use prompt engineering to instruct the model

Question 7

A financial services company wants to use generative AI to generate personalized investment advice. They must ensure responses comply with regulatory requirements (e.g., no guarantees of returns). Which Vertex AI safety feature should they primarily use?

Accepted Answer

Safety filters with a custom blocklist that includes phrases like 'guaranteed return'.. Safety filters with a custom blocklist allow the company to define specific prohibited phrases (e.g., 'guaranteed return') that the model must avoid generating. This provides a deterministic, rule-based enforcement layer that directly addresses regulatory compliance by blocking disallowed content at inference time, without relying on the model's probabilistic behavior.

Answer

Vertex AI Grounding with their compliance database.

Answer

Prompt engineering with instructions to avoid guarantees.

Answer

Reinforcement learning from human feedback (RLHF) on the model.

Question 8

A legal firm wants to automate contract analysis to extract key clauses and risks. They have 10,000 contracts in PDF format. The solution must handle varying layouts and be cost-effective. Which approach is BEST?

Accepted Answer

Use Document AI to convert PDFs to structured text, then use Vertex AI with a prompt that specifies clauses to extract. Using Document AI to parse PDFs into text, then Vertex AI with a structured prompt for clause extraction combines robust document understanding with flexible GenAI. Fine-tuning on 100 contracts is insufficient for layout variation. Agent Builder is overkill. Direct Gemini API on raw PDFs loses document structure.

Answer

Build a retrieval-augmented generation (RAG) system in Vertex AI Agent Builder

Answer

Use the Gemini API directly with the raw PDF files as input

Answer

Fine-tune a Gemini model on 100 annotated contracts and run inference on all contracts

Question 9

A research team is training a large language model from scratch using TPUs on Google Cloud. Which storage solution provides the highest throughput for training data?

Accepted Answer

Cloud Storage. Cloud Storage provides the highest throughput for training data because it is designed for high-bandwidth, parallel access from TPU pods via the Google Cloud Storage FUSE or gRPC-based data loading. TPUs benefit from Cloud Storage's ability to serve data at hundreds of GB/s when using the `tf.data` service with `tf.io.gfile` or the `gcloud storage` API, avoiding the I/O bottlenecks of block storage. Persistent Disk and Filestore have lower aggregate throughput limits and are not optimized for the distributed, streaming read patterns typical of large-scale training.

Answer

Persistent Disk

Answer

Cloud Filestore

Answer

Cloud Spanner

Question 10

An organization uses an IAM policy for Vertex AI as shown. A security audit reveals that engineer@example.com deployed a model that inadvertently exposed sensitive data. What is the most likely reason this happened?

Accepted Answer

The engineer had the aiplatform.user role, which includes permissions to deploy models without additional review.. The `aiplatform.user` role in Vertex AI includes the `aiplatform.model.deploy` permission, which allows any user with that role to deploy models without requiring additional approvals or administrative review. This lack of a secondary authorization step means the engineer could deploy a model that exposed sensitive data, even if the model had not been properly vetted for data leakage.

Answer

Audit logging is not enabled for DATA_WRITE events.

Answer

The admin user did not review the deployment.

Answer

The policy does not include a separation of duties between development and production.

Question 11

Which THREE are valid methods to reduce bias in generative AI outputs?

Accepted Answer

Using a more diverse training dataset. Training on a more diverse dataset reduces representational bias by exposing the model to a wider range of demographics, cultures, and perspectives. This directly mitigates the model's tendency to overrepresent majority groups or underrepresent minorities, which is a root cause of biased outputs in generative AI.

Answer

Using only English prompts

Answer

Increasing model size

Question 12

A company is deploying a generative AI application that generates medical reports. They need to ensure the output is factual and minimizes hallucinations. Which approach is most effective?

Accepted Answer

Implement retrieval-augmented generation (RAG) with a curated knowledge base. Retrieval-Augmented Generation (RAG) is the most effective approach because it grounds the model's output in a curated, authoritative knowledge base of medical data. By retrieving relevant, verified documents at inference time, RAG directly reduces the model's reliance on its parametric memory, which is the primary source of hallucinations in generative AI. This is especially critical in high-stakes domains like medical reporting, where factual accuracy is paramount.

Answer

Fine-tune the model with RLHF

Answer

Set the temperature to 0.0

Answer

Use prompt engineering to instruct the model to be accurate

Question 13

A data scientist is fine-tuning a large language model for a specialized domain using limited labeled data. To avoid catastrophic forgetting and reduce computational cost, which approach is recommended?

Accepted Answer

Adapter-based fine-tuning using LoRA. Adapter-based fine-tuning using LoRA (Low-Rank Adaptation) is recommended because it freezes the pre-trained model weights and injects trainable low-rank matrices into the transformer layers. This approach drastically reduces the number of parameters to update (often by 10,000x), lowering memory and compute requirements, while preserving the original knowledge to prevent catastrophic forgetting on limited domain data.

Answer

Using prompt engineering with in-context learning

Answer

Full fine-tuning of all model parameters

Answer

Training a new model from scratch on the domain data

Question 14

A company uses Gemini for Google Workspace to help employees draft emails. After rollout, the IT team notices that the Smart Compose suggestions are sometimes inappropriate for external communications. What is the BEST way to address this while maintaining productivity gains?

Accepted Answer

Train employees to write more specific prompts and provide few-shot examples for desired tone. Providing few-shot examples or using custom prompts for tone can guide the model to produce more appropriate suggestions. Disabling the feature or using a generic model would not leverage Workspace integration.

Answer

Switch to a third-party email assistant not integrated with Workspace

Answer

Disable Smart Compose for all external email accounts

Answer

Fine-tune the Gemini model on the company's past external emails

Question 15

A healthcare startup wants to generate synthetic patient notes for training medical residents. They need the output to follow a strict template with sections: Chief Complaint, History, Assessment, Plan. Which prompt engineering strategy should they use to ensure consistent structure?

Accepted Answer

Provide a few-shot example of the template filled out and instruct the model to follow that format for new cases. Structured output via few-shot examples showing the exact template ensures the model learns the required format. Adding instructions alone may not enforce the structure as reliably.

Answer

Use a zero-shot prompt asking for a patient note in bullet points

Answer

Set temperature to 0 and max tokens to a high value

Answer

Use a system prompt that lists the sections as instructions

Question 16

A developer deployed a large language model on Vertex AI for real-time chat. Users report slow response times. The model generates sentences one word at a time. Which optimization should be applied to reduce latency?

Accepted Answer

Use streaming responses to start output earlier.. Streaming responses allow the model to send tokens to the client as they are generated, rather than waiting for the full sequence to complete. This reduces perceived latency significantly in real-time chat, as users see the first word appear almost immediately, even though the total generation time remains similar.

Answer

Batch multiple user queries together.

Answer

Deploy the model with more accelerators.

Answer

Enable prompt caching to reuse previous queries.

Question 17

A sales team wants to use Gemini for Google Workspace to draft emails in Gmail. Which feature should they enable to generate full email drafts from a brief prompt?

Accepted Answer

Help me write. 'Help me write' is the specific Gemini for Google Workspace feature designed to generate full email drafts from a brief prompt. It leverages the underlying large language model to compose complete, context-aware email content, unlike features that only suggest short phrases or replies.

Answer

Translate

Answer

Smart Reply

Answer

Smart Compose

Question 18

A developer is using Vertex AI Gemini API for a chatbot. The chatbot sometimes outputs harmful content. What is the best first step to mitigate this?

Accepted Answer

Use safety filters and safety settings in the API request. The Vertex AI Gemini API provides built-in safety filters and configurable safety settings (e.g., `safety_settings` parameter with categories like `HARM_CATEGORY_HARASSMENT` and thresholds like `BLOCK_ONLY_HIGH`) that allow developers to block harmful outputs at inference time without retraining. This is the fastest and most direct first step to mitigate harmful content, as it requires no additional infrastructure or model modification.

Answer

Fine-tune the model on curated safe data

Answer

Add a human-in-the-loop review

Answer

Switch to a smaller model

Question 19

Which TWO techniques are effective for reducing bias in generative AI model outputs?

Accepted Answer

Training on diverse and representative datasets. Training on diverse and representative datasets directly reduces sampling bias and coverage gaps in the training distribution, which are primary sources of stereotypical or skewed outputs. By ensuring the model sees balanced examples across demographics, contexts, and edge cases, it learns more equitable representations and reduces the likelihood of generating biased content.

Answer

Increasing model size to learn more patterns

Answer

Relying solely on post-hoc filters

Answer

Limiting the model to only factual prompts

Question 20

According to Google's AI Principles, AI systems should be 'accountable to people.' What does this principle primarily require?

Accepted Answer

Organizations should provide avenues for feedback and recourse when AI systems cause harm. Accountability means providing mechanisms for feedback and recourse when AI systems cause harm or make mistakes.

Answer

AI systems should always have a human monitor present during operation

Answer

The AI system should be able to explain all its decisions to end users

Answer

AI systems should be designed to maximize shareholder value