Knowledge + Practice

Google Cloud Generative AI Leader Generative AI Leader (Generative AI Leader) — Questions 601–675

683 questions total · 10pages · All types, answers revealed

Take a mock exam Exam hub

Page 9 of 10

601

Multi-Selecthard

A company is deploying a GenAI system that generates product descriptions. During A/B testing, the new system shows a 20% increase in click-through rate (CTR) but a 15% increase in average cost per query due to the model size. The team wants to optimize cost without sacrificing the CTR gain. Which THREE actions should they take? (Choose three.)

Select 3 answers

A.Batch similar requests to reduce per-request overhead

B.Use a larger model with higher accuracy to further increase CTR

C.Increase the number of few-shot examples in the prompt

D.Switch to a smaller model and re-A/B test to confirm CTR impact

E.Implement response caching for repeated product SKUs

AnswersA, D, E

Batching reduces the number of API calls and can lower cost.

Why this answer

Batching similar requests reduces the per-request overhead by combining multiple inference calls into a single batch, which amortizes the fixed costs (e.g., model loading, token processing) across more outputs. This directly lowers the average cost per query while preserving the model architecture and CTR gains, as the model's output quality remains unchanged.

Exam trap

Google often tests the misconception that adding more few-shot examples always improves output quality, but in reality, it increases token costs and can degrade performance due to context window limits or irrelevant examples.

Full explanation →

602

MCQmedium

A research team wants to use Google's AI to generate video content from text prompts for a creative project. Which Google Cloud generative AI model should they use?

A.Imagen

B.Codey

C.Veo

D.Gemini

AnswerC

Veo is a generative video model that can create videos from text and image prompts.

Why this answer

Veo is Google's video generation model. Imagen generates images, Gemini is multimodal but not primarily for video generation, and Codey is for code.

Full explanation →

603

Multi-Selecthard

Which THREE factors should be considered when choosing between fine-tuning and prompt engineering for a generative AI task? (Choose three.)

Select 3 answers

A.Availability of labeled training data

B.Cost of API calls per request

C.Latency requirements for the application

D.Degree of task specialization required

E.Size of the base model

AnswersA, C, D

Fine-tuning needs labeled data.

Why this answer

Fine-tuning requires a labeled dataset specific to the target task to adjust model weights via supervised learning, whereas prompt engineering relies on the model's existing knowledge without additional training data. Without sufficient labeled data, prompt engineering is often the only viable approach, as fine-tuning would risk overfitting or poor generalization.

Exam trap

Google Cloud often tests the misconception that cost or model size are primary decision factors, when in reality the core trade-off is between data availability (labeled vs. unlabeled) and the degree of task specialization required.

Full explanation →

604

MCQeasy

Which Google Cloud service allows you to run machine learning models directly using SQL queries on data in BigQuery?

A.Cloud Functions

B.BigQuery ML

C.Vertex AI

D.Dataflow

AnswerB

BigQuery ML allows SQL-based ML directly in BigQuery.

Why this answer

BigQuery ML enables users to create, train, and deploy ML models using standard SQL, eliminating the need to move data to a separate environment.

Full explanation →

605

Multi-Selectmedium

A data scientist is fine-tuning a generative AI model for customer sentiment analysis. To ensure the fine-tuned model does not inadvertently memorize and reproduce personally identifiable information (PII) from the training data, which THREE practices should they follow? (Select 3)

Select 3 answers

A.Apply differential privacy during fine-tuning

B.Apply model quantization to reduce model size

C.Remove or anonymize all PII from the training data

D.Use only a subset of data that is necessary for the task (data minimization)

E.Use k-fold cross-validation to evaluate the model

AnswersA, C, D

Differential privacy bounds the influence of any single data point.

Why this answer

Differential privacy limits memorization; PII removal prevents it from being present; data minimization reduces risk. Quantization and k-fold cross-validation do not directly address PII memorization.

Full explanation →

606

Multi-Selecteasy

Which TWO are key business considerations when adopting generative AI solutions?

Select 2 answers

A.Training duration on public datasets

B.Number of model parameters

C.Data privacy and compliance requirements

D.Model accuracy on benchmarks

E.Cost of inference per request

AnswersC, E

Privacy and compliance are critical business and legal considerations.

Why this answer

Data privacy and compliance requirements (Option C) are a key business consideration because generative AI models often process sensitive or proprietary data, and regulations like GDPR, HIPAA, or CCPA mandate strict controls on data handling, storage, and model training. Failure to address these can result in legal penalties, reputational damage, and loss of customer trust, making it a top priority for enterprise adoption.

Exam trap

Google Cloud often tests the distinction between technical metrics (like training duration, parameter count, and benchmark accuracy) and true business considerations (like compliance, cost, and scalability), leading candidates to confuse model performance indicators with strategic business drivers.

Full explanation →

607

MCQhard

An enterprise deploys a generative AI chatbot that must comply with GDPR right to deletion. Users can request deletion of their personal data. The chatbot uses a RAG pipeline with a vector database. What is the MOST effective way to handle deletion requests?

A.Delete the user's documents from the vector index and original storage, then rebuild the index

B.Update the user's records in the vector index with anonymized placeholders

C.Add a filter to the chat application to block the user's name from appearing in responses

D.Retrain the LLM from scratch without the user's data

AnswerA

This removes the data from retrieval so the chatbot cannot access it, satisfying GDPR deletion requirements.

Why this answer

GDPR requires that personal data be erased when requested. In a RAG system, the source documents must be deleted from the vector store and the original storage. Fine-tuning or model training is not the right approach.

Full explanation →

608

Multi-Selectmedium

A company wants to build a GenAI-powered customer support chatbot. They require the chatbot to provide accurate answers based on the latest product documentation, and they need to control costs by minimizing token usage. Which TWO strategies should they use?

Select 2 answers

A.Use the largest available model for maximum accuracy

B.Cache frequent queries and their responses

C.Implement Retrieval-Augmented Generation (RAG) to retrieve relevant document chunks

D.Use a longer context window to include entire documents in the prompt

E.Fine-tune a large language model on the product documentation

AnswersB, C

Caching reduces repeated API calls, saving tokens and cost.

Why this answer

RAG ensures answers are grounded in latest docs. Caching common queries reduces token usage. Fine-tuning is expensive and not dynamic.

Using a large model increases cost. Long context windows increase token usage.

Full explanation →

609

MCQhard

A company is using Vertex AI to generate personalized marketing emails. The model sometimes produces biased content. What is the most effective way to detect and mitigate bias?

A.Add more diverse training data

B.Manually review all generated emails before sending

C.Switch to a different generative model

D.Use Vertex AI Explainable AI to analyze predictions and detect bias in training data

AnswerD

Explainable AI helps identify bias sources.

Why this answer

Vertex AI Explainable AI provides feature attributions and model explanations that help identify which input features (e.g., demographic attributes, phrasing patterns) contribute most to biased outputs. By analyzing these attributions against training data, you can pinpoint and mitigate bias at the source, rather than relying on post-hoc manual review or model swapping.

Exam trap

Google Cloud often tests the misconception that bias mitigation is solely a data quantity problem, leading candidates to choose 'add more diverse training data' without recognizing the need for diagnostic tools like Explainable AI to first detect and understand the bias.

How to eliminate wrong answers

Option A is wrong because simply adding more diverse training data does not guarantee detection of existing bias; it may reduce bias over time but lacks the diagnostic capability to identify specific biased patterns. Option B is wrong because manual review of all generated emails is not scalable, introduces human bias, and does not address the root cause of bias in the model's training or architecture. Option C is wrong because switching to a different generative model does not inherently detect or mitigate bias; the new model may have similar or different biases, and the underlying issue of biased training data or model behavior remains unaddressed.

Full explanation →

610

Multi-Selectmedium

A company is deploying a generative AI system for medical diagnosis. Which TWO measures are essential for responsible AI in this high-stakes domain?

Select 2 answers

A.Allow the AI to make autonomous decisions in time-sensitive emergencies

B.Ensure a human medical professional reviews all AI-generated diagnoses

C.Publish all patient data used for training to ensure transparency

D.Provide model documentation (Model Cards) to clinicians detailing the system's limitations

E.Use the AI only for administrative tasks, not diagnosis

AnswersB, D

Human oversight is critical for high-stakes decisions to catch errors and maintain accountability.

Why this answer

High-stakes domains require human oversight and transparency. Human review ensures accountability, and Model Cards communicate capabilities and limitations to stakeholders.

Full explanation →

611

MCQeasy

You are a generative AI lead at a healthcare startup developing a system to summarize patient medical records for quick review by doctors. The system uses a fine-tuned LLM. After deployment, doctors report that the summaries often miss critical details like medication dosages and allergy information. The current pipeline preprocesses patient records by extracting text from EHR, feeding it to the LLM, and outputting a summary. The team has limited time and budget. They cannot retrain the model because it is hosted as a managed API. Which action should you take to most effectively improve the summarization quality without changing the model?

A.Increase the maximum output token limit to force the model to include more details.

B.Replace the LLM with a simpler extractive summarization model that selects sentences from the original document.

C.Implement a retrieval-augmented generation (RAG) system that pulls supplementary data from external drug databases.

D.Revise the prompt to explicitly ask for medication dosages and allergies, and format the input text by adding headings (e.g., '### Medications') to emphasize important sections.

AnswerD

Prompt engineering is a low-cost, no-model-change solution that can emphasize key information.

Why this answer

Prompt engineering is the most effective and cost-efficient way to improve LLM output without retraining or changing the model. By explicitly instructing the model to include medication dosages and allergies, and by structuring the input with clear headings, you guide the model's attention to critical sections, directly addressing the missing details. This approach leverages the LLM's existing capabilities and requires no changes to the hosted API or additional infrastructure.

Exam trap

Google Cloud often tests the misconception that increasing output length or adding external data automatically improves quality, when in fact the most direct and cost-effective fix is to refine the input prompt to guide the model's focus.

How to eliminate wrong answers

Option A is wrong because increasing the maximum output token limit does not force the model to include specific missing details; it only allows longer responses, which may still omit critical information if the prompt does not direct the model's focus. Option B is wrong because replacing the LLM with an extractive summarization model would require retraining or deploying a new model, contradicting the constraint of not changing the model, and extractive methods cannot generate new text to explicitly mention dosages or allergies if they are not present in the original text. Option C is wrong because implementing a RAG system to pull from external drug databases adds complexity, cost, and latency, and does not address the core issue of missing details from the patient's own records; the problem is about extracting existing information, not supplementing with external data.

Full explanation →

612

Multi-Selecthard

A company is fine-tuning a Gemma model using Vertex AI. They observe that the model overfits. Which TWO actions should they take to mitigate overfitting?

Select 2 answers

A.Use a larger batch size

B.Increase the number of training epochs

C.Use more diverse data

D.Reduce the learning rate

E.Add dropout during fine-tuning

AnswersC, E

More diverse training data reduces overfitting to narrow patterns.

Why this answer

Introducing more diverse data helps the model generalize better by exposing it to a wider variety of patterns, reducing the risk of memorizing noise from a limited dataset. Option E is correct because dropout randomly deactivates a fraction of neurons during fine-tuning, which prevents co-adaptation and acts as a regularization technique to combat overfitting in transformer-based models like Gemma.

Exam trap

Google Cloud often tests the misconception that reducing the learning rate or increasing batch size are universal fixes for overfitting, when in fact these hyperparameters primarily affect optimization dynamics rather than regularization.

Full explanation →

613

MCQeasy

A retail company wants to integrate generative AI into its customer service chatbot to handle routine inquiries. They have a limited budget and want to launch quickly. Which strategy is most appropriate?

A.Partner with a generative AI vendor for a custom solution

B.Use pre-trained models via Google Cloud's Generative AI Studio API

C.Fine-tune an open-source model on their customer service logs

D.Build a custom LLM from scratch using the company's own data

AnswerB

Using pre-trained models via API is cost-effective and fast to implement.

Why this answer

Using pre-trained models via Google Cloud's Generative AI Studio API allows the company to leverage existing, powerful models without the high cost and time investment of custom development or fine-tuning. This approach enables rapid deployment on a limited budget by simply integrating the API into their chatbot, handling routine inquiries effectively without requiring extensive machine learning expertise or infrastructure.

Exam trap

Google Cloud often tests the misconception that fine-tuning or custom models are always better for domain-specific tasks, but the trap here is that for routine inquiries with limited budget and time, pre-trained APIs offer the fastest and most cost-effective solution without sacrificing quality.

How to eliminate wrong answers

Option A is wrong because partnering with a generative AI vendor for a custom solution typically involves significant upfront costs, long development cycles, and vendor lock-in, which contradicts the company's limited budget and need for quick launch. Option C is wrong because fine-tuning an open-source model on customer service logs requires substantial computational resources, data preparation, and machine learning expertise, making it slower and more expensive than using a pre-trained API. Option D is wrong because building a custom LLM from scratch is extremely resource-intensive, requiring massive datasets, specialized hardware, and months of training, which is impractical for a company with limited budget and a need for speed.

Full explanation →

614

Multi-Selecteasy

A developer is using Vertex AI Studio to test a text generation model. Which two actions can be performed in Vertex AI Studio? (Choose TWO)

Select 2 answers

A.Manage IAM roles

B.Monitor model cost

C.Create a dataset

D.Deploy a model to an endpoint

E.Fine-tune a model

AnswersD, E

From Studio, you can deploy a fine-tuned model directly to an endpoint.

Why this answer

Vertex AI Studio provides a direct interface to deploy a text generation model to an endpoint for serving predictions. This action is a core capability of the platform, allowing developers to test and then operationalize their models without leaving the Studio environment.

Exam trap

Google Cloud often tests the distinction between actions performed within a specific tool (Vertex AI Studio) versus broader platform capabilities (IAM, cost monitoring, dataset creation) to see if candidates understand the scope and purpose of each service.

Full explanation →

615

MCQhard

A company is deploying a chatbot using Vertex AI and wants to ensure that the model's responses are grounded in Google Search results to reduce hallucinations. Which feature should they enable?

A.Vertex AI Agent Builder

B.Retrieval-Augmented Generation (RAG) with internal documents

C.Gemini API with custom fine-tuning

D.Vertex AI Search for grounding

AnswerD

Vertex AI provides Google Search grounding to use search results as a knowledge source.

Why this answer

Google Search grounding in Vertex AI allows the model to retrieve real-time information from Google Search to ground answers. RAG uses private data, not web search. Vertex AI Agent Builder includes grounding but is for building agents, not specifically the grounding feature itself.

Full explanation →

616

MCQeasy

An e-commerce company is using a generative AI model to recommend products. They notice that the recommendations are often irrelevant. What is the most likely cause?

A.Using an outdated model version

B.Incorrect regional endpoint configuration

C.Inadequate prompt engineering

D.Overfitting on training data

AnswerC

The model's output quality heavily depends on the prompt; poor prompts lead to irrelevant responses.

Why this answer

Inadequate prompt engineering is the most likely cause because generative AI models rely heavily on the quality and specificity of the input prompt to produce relevant outputs. If the prompts used to generate product recommendations are vague, poorly structured, or lack context (e.g., not including user preferences or historical behavior), the model will return generic or irrelevant suggestions. This is a common failure point in recommendation systems where the prompt acts as the primary interface for steering model behavior.

Exam trap

Google Cloud often tests the misconception that model performance issues are always due to training data or model version problems, when in fact prompt engineering is the most immediate and common cause of output irrelevance in generative AI systems.

How to eliminate wrong answers

Option A is wrong because using an outdated model version may affect performance or feature availability, but it does not directly cause irrelevant recommendations; the model would still generate outputs consistent with its training, and relevance is more tied to prompt quality. Option B is wrong because incorrect regional endpoint configuration would cause connectivity or latency issues (e.g., API timeouts or routing errors), not irrelevant content generation; the model's output relevance is independent of the endpoint's geographic location. Option D is wrong because overfitting on training data would cause the model to memorize specific patterns and perform poorly on new or diverse inputs, but in a recommendation context, overfitting typically leads to overly narrow or repetitive suggestions, not broadly irrelevant ones; the primary issue with irrelevant outputs is prompt misalignment, not training data memorization.

Full explanation →

617

MCQmedium

A financial services firm is using a foundation model on Vertex AI to generate investment summaries from quarterly reports. The summaries are accurate but often miss key financial metrics and trends. The team cannot afford to fine-tune the model frequently. Which technique should they use to improve the completeness and relevance of the summaries without modifying the model?

A.Increase temperature to 0.9 to encourage more creative outputs.

B.Provide three few-shot examples in the prompt that highlight the desired metrics.

C.Set stop sequences to [' '] to ensure the model finishes each paragraph.

D.Lower top_p to 0.5 to reduce the sampling pool.

AnswerB

Few-shot examples condition the model to replicate the structure and content of the examples.

Why this answer

Few-shot prompting provides the model with concrete examples of desired output structure and content, guiding it to include key financial metrics and trends without retraining. This technique leverages in-context learning, where the model generalizes from the examples in the prompt to produce more complete and relevant summaries, while avoiding the cost and latency of fine-tuning.

Exam trap

The trap here is that candidates confuse hyperparameter tuning (temperature, top_p) with prompt engineering, assuming that increasing randomness or restricting token selection will improve output quality, when in fact few-shot examples directly teach the model the desired output structure without modifying the model.

How to eliminate wrong answers

Option A is wrong because increasing temperature to 0.9 encourages randomness and creativity, which would likely make summaries less focused and more prone to missing key metrics, not more complete. Option C is wrong because setting stop sequences to ['

'] only controls when the model stops generating text, but does not influence the content or inclusion of specific financial metrics within the output. Option D is wrong because lowering top_p to 0.5 reduces the sampling pool to only the most likely tokens, which can make outputs more repetitive and less likely to include diverse or specific metrics, not improve completeness.

Full explanation →

618

MCQmedium

A developer is using the Vertex AI Gemini API to generate product descriptions. They get a 400 error 'INVALID_ARGUMENT: The model's maximum input token limit is 8192.' What is the most likely issue?

A.The prompt is too long

B.The API key is invalid

C.The output tokens are too high

D.The model is not available in the region

AnswerA

The error explicitly states the input token limit is exceeded.

Why this answer

The 400 error 'INVALID_ARGUMENT: The model's maximum input token limit is 8192' explicitly indicates that the combined token count of the prompt (system instructions, user input, and any conversation history) exceeds the 8192-token context window of the Gemini model being used. This is a hard limit enforced by the Vertex AI Gemini API, and the error is triggered before any generation begins. Therefore, the most likely issue is that the prompt is too long.

Exam trap

The trap here is that candidates confuse input token limits with output token limits or general API authentication errors, but the specific error message 'maximum input token limit' directly points to prompt length as the root cause.

How to eliminate wrong answers

Option B is wrong because an invalid API key would result in a 401 Unauthorized or 403 Forbidden error, not a 400 INVALID_ARGUMENT error related to token limits. Option C is wrong because the error message specifically mentions 'input token limit', not output tokens; output token limits are enforced separately (e.g., via max_output_tokens parameter) and would produce a different error. Option D is wrong because model availability in a region would cause a 404 or 403 error (e.g., 'Model not found' or 'Permission denied'), not a token-limit-related INVALID_ARGUMENT error.

Full explanation →

619

MCQmedium

A healthcare organization needs to run ML models on patient data stored in BigQuery while ensuring data never leaves the database. Which service allows them to create and execute ML models directly in BigQuery SQL?

A.Cloud Functions

B.Vertex AI Prediction

C.Vertex AI Workbench

D.BigQuery ML

AnswerD

BigQuery ML allows users to create, train, and evaluate ML models using SQL queries directly on BigQuery data.

Why this answer

BigQuery ML enables SQL-based ML model creation and execution directly on data in BigQuery, meeting the requirement without data movement.

Full explanation →

620

MCQeasy

A company is choosing between Google's Gemini API and an open-source model. Which factor is most important for a business with limited ML expertise?

A.Ease of integration and availability of support

B.Model parameter count

C.Cost per token

D.Community size

AnswerA

Limited ML expertise means the team needs a solution that is easy to integrate and comes with reliable support.

Why this answer

For a business with limited ML expertise, ease of integration and availability of support are paramount because they reduce the need for in-house machine learning engineering talent. Google's Gemini API offers managed infrastructure, pre-built SDKs, and enterprise-grade support (e.g., SLA-backed uptime, dedicated account management), which directly lowers the barrier to entry and operational risk. In contrast, open-source models require significant expertise for deployment, scaling, and troubleshooting, making them unsuitable for teams without deep ML skills.

Exam trap

The Generative AI Leader exam often tests the misconception that technical metrics like parameter count or cost per token are the primary decision factors, when in reality, for a non-expert team, operational simplicity and vendor support are the critical success factors that determine whether a GenAI project can be delivered at all.

How to eliminate wrong answers

Option B is wrong because model parameter count (e.g., 7B vs 175B) is a technical metric that does not directly address the business's lack of ML expertise; a larger parameter count can actually increase complexity and resource requirements, making it harder to integrate without expert knowledge. Option C is wrong because cost per token, while important for budgeting, is secondary to the ability to actually use the model; without easy integration and support, even a low-cost model can become expensive due to hidden engineering costs and downtime. Option D is wrong because community size, though helpful for troubleshooting, does not provide the structured, guaranteed support and SLAs that a business with limited ML expertise needs; community forums lack accountability and may not offer timely or accurate solutions for production-critical issues.

Full explanation →

621

MCQhard

What is the most likely cause of the error?

A.The predict schema must be stored in the same bucket as the model artifacts and referenced without the full gs:// URI

B.The display name contains a hyphen which is not allowed

C.The container image URI is incorrect

D.The region us-central1 does not support TensorFlow models

AnswerA

The schema should be a relative path within the artifact URI.

Why this answer

The error occurs because the Vertex AI Predict schema must be stored in the same Cloud Storage bucket as the model artifacts, and when referenced in the model upload request, it should use a relative path (without the full `gs://` URI). Using the full URI causes a parsing failure, as Vertex AI expects the schema to be co-located with the model artifacts for validation and deployment.

Exam trap

Google Cloud often tests the nuance that Vertex AI expects schema files to be co-located with model artifacts and referenced without the full `gs://` URI, causing candidates to incorrectly assume the error is due to region limitations or container image issues.

How to eliminate wrong answers

Option B is wrong because hyphens are allowed in display names for Vertex AI models; the constraint is on the model ID (auto-generated) and display names can contain hyphens, underscores, and alphanumeric characters. Option C is wrong because the container image URI is syntactically correct and points to a valid Vertex AI pre-built serving image for TensorFlow; the error is not related to the container URI format. Option D is wrong because us-central1 fully supports TensorFlow models on Vertex AI; it is one of the primary regions for AI Platform and Vertex AI model deployment.

Full explanation →

622

MCQeasy

Which Google Cloud AI service provides a unified ML platform for building, deploying, and managing ML models in production?

A.AI Platform

B.Vertex AI

C.BigQuery ML

D.Cloud AutoML

AnswerB

Vertex AI is the single platform for all ML activities, including training, deployment, and management.

Why this answer

Vertex AI is the correct answer because it is Google Cloud's unified ML platform that integrates data engineering, data science, and ML engineering workflows into a single service. It provides end-to-end capabilities for building, training, deploying, and managing ML models in production, including AutoML, custom training, model registry, and MLOps features like continuous evaluation and monitoring.

Exam trap

The trap here is that candidates often confuse AI Platform (the legacy service) with Vertex AI, not realizing that Vertex AI is the successor that consolidates all ML capabilities into a single, unified platform, making AI Platform a deprecated option in the context of current Google Cloud ML strategy.

How to eliminate wrong answers

Option A is wrong because AI Platform (now legacy) was the predecessor to Vertex AI; it lacked the unified integration of AutoML and custom training under a single API and did not provide the same level of MLOps tooling, such as model monitoring and feature store. Option C is wrong because BigQuery ML is a service that allows users to create and execute ML models using SQL queries directly in BigQuery, but it is not a unified ML platform for building, deploying, and managing models in production—it is limited to in-database ML and does not support custom training frameworks or production deployment pipelines. Option D is wrong because Cloud AutoML is a subset of Vertex AI that focuses on training high-quality models with minimal effort using Google's transfer learning and neural architecture search, but it does not provide the full platform capabilities for custom model development, deployment, and management that Vertex AI offers.

Full explanation →

623

MCQhard

A healthcare company is using a fine-tuned version of PaLM 2 on Vertex AI to generate clinical notes from doctor-patient conversations. The model was fine-tuned on a dataset of 10,000 de-identified transcripts and corresponding notes. During testing, the generated notes are grammatically correct and well-structured, but they often contain subtle inaccuracies: for example, they might mention a medication that was not discussed, or omit a key symptom. The team has already tried increasing the training epochs and adjusting learning rates, with minimal improvement. They need a solution that can be implemented quickly to improve factual accuracy without retraining the entire model. The team has access to a large archive of verified clinical notes and a small set of recent conversation-to-note pairs that have been manually reviewed and corrected. The inference pipeline currently uses a single call to the model with the conversation transcript as input. What should the team do?

A.Implement retrieval-augmented generation (RAG) by retrieving similar verified notes from the archive and providing them as context in the prompt.

B.Decrease the temperature to 0.1 to reduce randomness and force the model to stick to the input.

C.Use prompt engineering to instruct the model to only include information explicitly mentioned in the conversation.

D.Add a human-in-the-loop step to review and correct every generated note before use.

AnswerA

RAG grounds the generation in factual examples, directly reducing inaccuracies without retraining.

Why this answer

Retrieval-augmented generation (RAG) directly addresses the core issue of factual inaccuracy without retraining. By retrieving verified clinical notes similar to the current conversation from the archive and injecting them as context in the prompt, the model gains access to ground-truth examples that anchor its output to factual details. This approach leverages the team's existing archive and small set of corrected pairs to provide relevant, accurate context, improving precision without modifying the model's weights.

Exam trap

The trap here is that candidates often assume factual inaccuracy is solely a randomness issue (temperature) or a prompt instruction problem, overlooking that the model's parametric knowledge is insufficient and needs external grounding via retrieval augmentation.

How to eliminate wrong answers

Option B is wrong because decreasing temperature to 0.1 reduces randomness but does not fix factual inaccuracies stemming from the model's training data or lack of context; it may actually cause the model to become overly deterministic and repeat hallucinations from its fine-tuning. Option C is wrong because prompt engineering to instruct the model to only include explicitly mentioned information is a superficial fix that cannot overcome the model's tendency to hallucinate or omit details when the training data or fine-tuning process has embedded those inaccuracies; it lacks the grounding provided by external verified data. Option D is wrong because adding a human-in-the-loop step to review every note is a manual, non-scalable solution that does not improve the model's output quality at inference time and fails to address the root cause of factual inaccuracy; it also contradicts the requirement for a quick implementation without retraining.

Full explanation →

624

MCQeasy

What is the primary purpose of Google's Datasheets for Datasets?

A.To serve as a legal contract for data sharing

B.To list all models trained on the dataset

C.To document the dataset's creation, composition, and intended use

D.To provide a template for labeling data

AnswerC

Datasheets provide a structured format for documenting datasets.

Why this answer

Datasheets for Datasets are designed to document the motivation, composition, collection process, and other details of a dataset to promote transparency and reproducibility.

Full explanation →

625

MCQhard

A hospital wants to summarize patient-doctor conversations into structured clinical notes using GenAI. They need high accuracy and must avoid hallucinated medical information. Which combination of techniques is BEST?

A.Use RAG with medical textbooks and a low temperature setting

B.Fine-tune a medical-specific model on de-identified transcripts and use a structured output format in the prompt

C.Use a large model with zero-shot prompting and post-process the output with rule-based checks

D.Use a Gemini model with a high temperature setting to encourage creativity

AnswerB

Fine-tuning on medical data improves accuracy; structured output reduces hallucination.

Why this answer

Fine-tuning with structured output (e.g., JSON schema) and a strict prompt ensures the model produces accurate, formatted notes and reduces hallucinations.

Full explanation →

626

MCQmedium

A healthcare startup wants to use Vertex AI to deploy a model that helps doctors diagnose rare diseases. The model must be explainable, showing the reasoning path. Which technique should they implement?

A.Use a black-box ensemble model for higher accuracy

B.Reduce the model size to make it inherently interpretable

C.Disable all safety filters to avoid interfering with model output

D.Implement chain-of-thought prompting to output reasoning steps

AnswerD

Chain-of-thought provides a clear reasoning path, aiding interpretability.

Why this answer

Chain-of-thought reasoning allows the model to generate step-by-step explanations, which is crucial for medical diagnosis transparency.

Full explanation →

627

MCQeasy

A company wants to build a text-to-speech application for generating voiceovers in multiple languages. They need to use a pre-built Google API without training custom models. Which service should they use?

A.Cloud Text-to-Speech API

B.Vertex AI Text-to-Speech with a custom model

C.Gemini API with a custom prompt

D.Cloud Speech-to-Text API

AnswerA

Cloud Text-to-Speech API provides pre-built voices in many languages, no custom training needed.

Why this answer

Cloud Text-to-Speech API is a pre-built service for converting text to natural-sounding speech in many languages. The other options are either for speech recognition or not pre-built APIs.

Full explanation →

628

MCQeasy

A startup wants to use a pre-trained model to generate product descriptions without training. Which Google Cloud service should they use?

A.Vertex AI Prediction

B.AI Platform Training

C.Cloud AutoML

D.Vertex AI Generative AI Studio

AnswerD

Vertex AI Generative AI Studio is designed for accessing and experimenting with foundation models for generative tasks.

Why this answer

Vertex AI Generative AI Studio is the correct service because it provides a no-code interface to access and experiment with pre-trained generative models, including text generation for product descriptions, without requiring any training or custom model development. It allows users to directly prompt models like PaLM 2 or Gemini for inference tasks, making it ideal for generating content from a pre-trained model without training.

Exam trap

The trap here is that candidates may confuse Vertex AI Prediction (which serves custom models) with Generative AI Studio (which serves pre-trained models), or assume that any generative AI task requires training via AI Platform Training or AutoML, when in fact the question explicitly states 'without training'.

How to eliminate wrong answers

Option A is wrong because Vertex AI Prediction is used for deploying and serving custom-trained models for online or batch predictions, not for directly accessing pre-trained generative models without training. Option B is wrong because AI Platform Training is designed for training custom machine learning models, not for using pre-trained models for inference without training. Option C is wrong because Cloud AutoML is used for training custom models on user-provided data with automated machine learning, not for directly generating content from a pre-trained model without any training.

Full explanation →

629

Multi-Selecthard

A financial services firm must comply with regulations when using gen AI. Which two measures are critical?

Select 2 answers

A.Implement audit trails

B.Deploy without risk assessment

C.Use a closed-source model

D.Use explainable AI

E.Use only synthetic data

AnswersA, D

Audit trails provide accountability and support regulatory reviews.

Why this answer

Audit trails are critical for compliance because they provide a tamper-evident, chronological record of all AI model inputs, outputs, and decisions. This enables firms to demonstrate regulatory adherence (e.g., under GDPR or SOX) by reconstructing the exact sequence of events that led to a specific AI-generated output, which is essential for accountability and forensic review.

Exam trap

Google Cloud often tests the misconception that 'closed-source models are inherently more compliant' or that 'synthetic data eliminates privacy risks,' when in reality, compliance hinges on transparency, auditability, and risk assessment rather than the model's source or data origin.

Full explanation →

630

MCQmedium

A data scientist is using Vertex AI Model-as-a-Service (MaaS) to deploy a fine-tuned open-source model. They notice high latency during inference. What is the most likely cause?

A.The model is too large for the hardware

B.The endpoint is set to autoscaling with a low minimum node count

C.The model is not quantized

D.The region is incorrect

AnswerB

Autoscaling with low min nodes causes cold start latency.

Why this answer

High latency during inference on Vertex AI MaaS is most often caused by the endpoint scaling configuration. When autoscaling is enabled with a low minimum node count, the system may need to provision additional nodes to handle the request load, which introduces cold-start latency. This is especially pronounced for fine-tuned open-source models, which can be large and take time to load onto new nodes.

Exam trap

The trap here is that candidates often assume high latency is due to model size or lack of optimization, but the question specifically describes 'high latency during inference' in a MaaS context, which points to scaling delays rather than compute bottlenecks.

How to eliminate wrong answers

Option A is wrong because Vertex AI MaaS automatically selects appropriate hardware for the model size; if the model were too large, it would fail to deploy rather than cause high latency. Option C is wrong because quantization primarily affects throughput and memory usage, not the latency spike from cold starts; unquantized models may be slower but not cause the intermittent high latency described. Option D is wrong because the region setting affects data residency and network latency, but it would cause consistently higher latency, not the intermittent high latency typical of autoscaling delays.

Full explanation →

631

MCQhard

A media company is using Vertex AI Imagen to generate marketing images. The output frequently contains unrealistic artifacts, especially in human faces. The team has fine-tuned the model using their brand assets. What is the most likely cause and recommended fix?

A.Safety filters are too aggressive; reduce them.

B.Negative prompts are missing; always include 'unrealistic'.

C.The fine-tuning dataset is too small or too homogeneous; augment and diversify the training data.

D.Inference steps are too low; increase to 100.

AnswerC

Overfitting to limited data causes artifacts; more varied data helps generalization.

Why this answer

Unrealistic artifacts in fine-tuned generative models, especially in human faces, typically stem from a training dataset that is too small or lacks diversity. When the dataset is homogeneous, the model overfits to limited patterns and fails to generalize, leading to distorted outputs. Augmenting and diversifying the training data with varied poses, lighting, and ethnicities helps the model learn robust facial features.

Exam trap

The trap here is that candidates confuse inference parameters (like steps or safety filters) with data quality issues, assuming artifacts are due to model settings rather than the fundamental cause of insufficient or non-diverse training data.

How to eliminate wrong answers

Option A is wrong because safety filters in Vertex AI Imagen block harmful content (e.g., violence, hate speech) and do not cause unrealistic artifacts; reducing them would not fix facial distortions and could introduce policy violations. Option B is wrong because negative prompts guide the model to avoid certain concepts, but simply including the word 'unrealistic' is not a technical fix—the model needs diverse training data, not a prompt hack. Option D is wrong because inference steps control the denoising process and image quality, but increasing them to 100 would not address overfitting from a poor dataset; the default steps (typically 50) are sufficient for high-quality outputs.

Full explanation →

632

MCQmedium

A healthcare organization wants to use generative AI for medical report summaries. What is the primary concern?

A.Ensuring HIPAA compliance and data security when using cloud AI services

B.The model's ability to generate fluent and coherent summaries

C.Minimizing the cost of each API call to stay within budget

D.Latency of responses for real-time use cases

AnswerA

Generative AI models processing PHI must be HIPAA-compliant, requiring a signed Business Associate Agreement (BAA) with Google Cloud.

Why this answer

The primary concern for a healthcare organization using generative AI for medical report summaries is ensuring HIPAA compliance and data security when using cloud AI services. Medical data is protected health information (PHI), and any cloud-based AI service must have a Business Associate Agreement (BAA) in place and enforce encryption at rest and in transit to avoid regulatory penalties and data breaches.

Exam trap

Google Cloud often tests the misconception that technical performance (fluency, cost, latency) is the top priority, when in regulated industries like healthcare, compliance and data security are the non-negotiable primary concerns.

How to eliminate wrong answers

Option B is wrong because while fluency and coherence are important for summary quality, they are secondary to the legal and security obligations of handling PHI; a fluent summary that leaks data is non-compliant. Option C is wrong because cost minimization is an operational concern, not the primary risk; HIPAA violations carry fines up to $50,000 per violation, far outweighing API call costs. Option D is wrong because latency is a performance metric relevant for real-time use, but medical report summarization is typically asynchronous or batch-processed, and compliance takes precedence over speed.

Full explanation →

633

Multi-Selectmedium

A company is using Vertex AI to build a language model for generating legal documents. They need to ensure the model's outputs are accurate and verifiable. Which TWO features should they use?

Select 2 answers

A.Confidence indicators

B.Safety filters for legal content

C.Chain-of-thought reasoning

D.Grounding with citations to relevant legal texts

E.Model Cards

AnswersC, D

Chain-of-thought makes the model's reasoning process transparent, aiding in verification.

Why this answer

Grounding (citing sources) and chain-of-thought reasoning are the two features that directly improve accuracy and verifiability. Confidence indicators and safety filters are useful but do not directly address accuracy/verifiability.

Full explanation →

634

MCQeasy

Which of the following best describes how large language models (LLMs) generate text?

A.They retrieve the most similar text from a database and return it

B.They use a rule-based grammar engine to construct sentences

C.They predict the next token in a sequence based on the preceding tokens

D.They randomly select words from a fixed vocabulary

AnswerC

LLMs use the transformer architecture to model the probability distribution of the next token given the context.

Why this answer

LLMs are trained to predict the next token given the preceding tokens. During inference, they generate one token at a time autoregressively.

Full explanation →

635

MCQhard

A developer receives the above JSON response from a Vertex AI language model. The output content is correct, but the developer expected the model to not answer geography questions. What should the developer do to prevent the model from responding to geography queries?

A.Adjust the safety filter thresholds for the 'Toxic' category

B.Enable Vertex AI Grounding with a geography knowledge base

C.Configure a safety filter for the 'Geography' category

D.Add a system instruction to not answer geography questions

AnswerD

Adding a system instruction to not answer geography questions is the correct method to enforce this custom restriction, as system instructions guide the model's behavior.

Why this answer

Vertex AI does not have a predefined safety filter category for 'Geography'. To prevent the model from answering geography questions, the developer should use a system instruction that explicitly tells the model not to respond to geography queries. System instructions are the appropriate mechanism for custom content restrictions, while safety filters are limited to predefined categories such as toxic, harassment, etc.

Exam trap

Candidates often assume safety filters can be configured for arbitrary topics like geography, but Vertex AI safety filters only support predefined categories. The correct approach is to use system instructions to guide model behavior for custom restrictions.

How to eliminate wrong answers

Option A is wrong because adjusting safety filter thresholds for the 'Toxic' category only controls responses related to toxicity (e.g., hate speech, harassment), not geography-specific content; it does not address the requirement to block geography questions. Option B is wrong because enabling Vertex AI Grounding with a geography knowledge base would actually enhance the model's ability to answer geography queries by providing additional context, which is the opposite of what the developer wants. Option D is wrong because while adding a system instruction to not answer geography questions might influence the model, it is not a guaranteed enforcement mechanism—models can still override or ignore instructions, especially if the prompt is rephrased; safety filters provide a more reliable, configurable block.

Full explanation →

636

MCQeasy

A healthcare company needs to process medical records (e.g., discharge summaries) to extract structured data. Which AI API is specifically designed for this purpose?

A.Natural Language AI

B.Translation AI

C.Document AI with Healthcare NLP

D.Vision AI

AnswerC

Healthcare AI (via Document AI Healthcare NLP) is specialized for medical records.

Why this answer

Healthcare AI is purpose-built for medical documents, providing features like entity extraction for clinical data.

Full explanation →

637

MCQhard

An MLOps engineer wants to implement continuous evaluation of a generative model in production. Which Vertex AI component should they use?

A.Vertex AI Model Monitoring

B.Vertex AI Feature Store

C.Vertex AI Prediction

D.Vertex AI Pipelines

AnswerA

Model Monitoring provides continuous evaluation of model metrics and alerts on degradation.

Why this answer

Vertex AI Model Monitoring is the correct component because it provides continuous evaluation of model performance in production, including detecting prediction drift, data drift, and feature attribution drift. For generative models, it can monitor output quality and safety metrics over time, alerting engineers to degradation or shifts in model behavior without requiring manual intervention.

Exam trap

Google Cloud often tests the distinction between monitoring (ongoing evaluation of deployed models) and serving (handling inference requests), leading candidates to mistakenly choose Vertex AI Prediction for continuous evaluation tasks.

How to eliminate wrong answers

Option B is wrong because Vertex AI Feature Store is designed for managing, storing, and serving feature data for training and predictions, not for monitoring model performance or evaluating outputs in production. Option C is wrong because Vertex AI Prediction handles model serving and inference requests, but it does not include built-in continuous evaluation or drift detection capabilities. Option D is wrong because Vertex AI Pipelines orchestrates ML workflows for training and batch prediction, but it is not a real-time monitoring service for production model evaluation.

Full explanation →

638

MCQhard

A music streaming service wants to use AI-generated playlists and artwork, but is concerned about potential copyright infringement. They plan to use a generative model that was trained on a large corpus of publicly available music and images. Which action is MOST important to mitigate IP risk?

A.Review the training data provenance and ensure it consists of properly licensed or public domain works

B.Only use models hosted on Google Cloud, as Google assumes liability

C.Add a watermark to all generated content using SynthID

D.Ask the model to self-certify that its outputs are original

AnswerA

Understanding and documenting the training data's legal status helps mitigate IP infringement risk.

Why this answer

Training data provenance is a key IP concern. The service should verify that the training data does not include copyrighted content without permission, and document the data sources to establish a chain of provenance.

Full explanation →

639

Multi-Selecteasy

Which THREE are essential components of a responsible AI strategy for GenAI? (Select three.)

Select 3 answers

A.Use of only open-source models

B.Maximum model size

C.Human oversight for critical decisions

D.Model transparency and explainability

E.Bias detection and mitigation

AnswersC, D, E

Human oversight prevents harmful automated decisions and ensures ethical use.

Why this answer

Human oversight for critical decisions (C) is essential because GenAI models can produce plausible but incorrect or harmful outputs. A responsible AI strategy mandates that a human-in-the-loop reviews high-stakes outputs, such as medical diagnoses or financial approvals, to prevent automated errors from causing real-world harm. This aligns with the principle of human accountability in AI governance frameworks like the NIST AI Risk Management Framework.

Exam trap

Google Cloud often tests the misconception that technical attributes like model size or open-source licensing are core to responsible AI, when in fact the focus is on governance practices like transparency, bias mitigation, and human oversight.

Full explanation →

640

MCQmedium

A startup is prototyping a multimodal AI application that processes images and text. They have a limited budget and want the fastest time to market, with minimal infrastructure setup. Which combination of services should they use for prototyping?

A.Vertex AI Prediction and Cloud Storage

B.Google AI Studio (Gemini API) and Colab

C.Cloud Run and Firestore

D.Vertex AI Workbench and BigQuery ML

AnswerB

Google AI Studio offers a free tier for Gemini API, and Colab provides free GPU notebooks — ideal for rapid prototyping with multimodal data.

Why this answer

Google AI Studio provides immediate access to the Gemini API for multimodal (image+text) processing without any infrastructure setup, and Colab offers a free, managed Jupyter environment with pre-installed libraries for rapid prototyping. This combination minimizes time to market and cost, aligning perfectly with the startup's constraints.

Exam trap

The trap here is that candidates often over-engineer the solution by choosing managed ML platforms like Vertex AI Prediction, forgetting that prototyping prioritizes speed and minimal setup over production-grade scalability.

How to eliminate wrong answers

Option A is wrong because Vertex AI Prediction requires deploying a model to an endpoint, which involves infrastructure setup and ongoing costs, making it slower and more expensive for prototyping. Option C is wrong because Cloud Run and Firestore are serverless compute and database services, not designed for multimodal AI processing; they would require building custom ML logic and lack built-in multimodal capabilities. Option D is wrong because Vertex AI Workbench is a managed notebook environment for model development, and BigQuery ML is for SQL-based ML on tabular data, neither providing direct multimodal AI inference like the Gemini API.

Full explanation →

641

MCQmedium

A team is using Vertex AI to fine-tune a large language model on proprietary company data. The data contains personally identifiable information (PII). What is the BEST practice to protect privacy?

A.Use differential privacy during the fine-tuning process

B.Use a different foundation model that was not trained on proprietary data

C.Remove all PII from the dataset before fine-tuning

D.Store the fine-tuned model on-premises only

AnswerA

Differential privacy adds noise to prevent the model from memorizing individual data points.

Why this answer

Differential privacy (Option A) is the best practice because it adds calibrated noise during fine-tuning, mathematically guaranteeing that the model cannot memorize or leak individual PII records even if the training data contains such information. This approach preserves privacy without requiring complete removal of PII, which may be impractical or destroy data utility. Vertex AI supports differential privacy through libraries like TensorFlow Privacy, enabling privacy budget tracking via epsilon (ε) values.

Exam trap

The Generative AI Leader exam often tests the misconception that data removal (Option C) is sufficient for privacy, when in fact differential privacy provides a formal mathematical guarantee against inference attacks even if PII is present in the training set.

How to eliminate wrong answers

Option B is wrong because using a different foundation model does not address the privacy risk; the fine-tuning process on proprietary data still exposes PII regardless of the base model chosen. Option C is wrong because removing all PII before fine-tuning is a data preprocessing step that can reduce risk but is not a privacy guarantee—residual PII may remain due to incomplete scrubbing, and the model can still infer sensitive patterns from non-PII fields. Option D is wrong because storing the fine-tuned model on-premises only addresses data residency but does not prevent the model from memorizing and leaking PII during inference; privacy protection requires algorithmic safeguards, not just storage location.

Full explanation →

642

MCQeasy

A company is deploying a generative AI model for medical advice. What is the most important consideration?

A.Model latency

B.Safety and fairness

C.Model size

D.Cost of inference

AnswerB

Patient safety and avoiding bias are the top priorities.

Why this answer

In medical advice applications, a generative AI model's outputs can directly impact patient health, making safety and fairness the paramount consideration. Incorrect or biased advice could lead to misdiagnosis or harm, outweighing performance metrics like latency or cost. Regulatory frameworks such as HIPAA and FDA guidelines for clinical decision support further mandate rigorous validation of model safety and fairness before deployment.

Exam trap

Google Cloud often tests the misconception that technical performance metrics like latency or cost are the primary concerns in high-stakes domains, when in fact ethical and safety considerations take precedence.

How to eliminate wrong answers

Option A is wrong because model latency, while important for user experience, is secondary to ensuring the advice is safe and unbiased; a fast but harmful response is unacceptable in healthcare. Option C is wrong because model size correlates with computational resources and potential capability, but does not inherently guarantee safety or fairness; a larger model may amplify biases or generate more confident but incorrect advice. Option D is wrong because cost of inference is a business consideration that must be balanced against safety requirements, but it is not the most critical factor when human lives are at stake.

Full explanation →

643

MCQmedium

A travel company fine-tuned a language model on customer chat logs to provide travel recommendations. After deployment, they receive complaints that the model sometimes generates inappropriate or offensive content. What is the most effective approach to improve output safety while preserving overall performance?

A.Modify the system instruction to request polite responses only

B.Retrain the model on a larger dataset of chat logs

C.Reduce the temperature to 0.0

D.Add a post-processing safety classifier that filters or rewrites unsafe outputs

AnswerD

A safety classifier directly catches and mitigates harmful content without modifying the base model.

Why this answer

A post-processing safety classifier acts as a guardrail that can detect and filter or rewrite unsafe outputs without altering the underlying model's weights or training data. This approach preserves the model's overall performance on safe, relevant recommendations while adding a dedicated safety layer that can be independently tuned and updated as new safety requirements emerge. Unlike prompt engineering or hyperparameter adjustments, a classifier provides a robust, policy-enforced mechanism to catch edge cases that the model might otherwise generate.

Exam trap

Google often tests the misconception that prompt engineering or hyperparameter tuning alone can reliably fix safety issues, when in fact they are insufficient against learned toxic patterns in the model's weights, and a dedicated safety classifier is the standard industry practice for robust output filtering.

How to eliminate wrong answers

Option A is wrong because modifying the system instruction is a form of prompt engineering that can be easily overridden by the model's learned patterns from training data; it does not guarantee the model will never generate offensive content, especially if such patterns exist in the fine-tuning data. Option B is wrong because retraining on a larger dataset of chat logs does not address the root cause of inappropriate outputs—if the original data contains toxic or biased examples, simply adding more data may dilute but not eliminate the problem, and could even introduce new unsafe patterns. Option C is wrong because reducing temperature to 0.0 makes the model deterministic and greedy, which reduces creativity but does not prevent the model from generating the most likely token sequence that could still be offensive; it also harms performance on diverse, nuanced travel recommendations.

Full explanation →

644

MCQeasy

A developer is using Vertex AI PaLM 2 to generate product descriptions. The output is often too verbose and includes irrelevant details. Which technique should the developer apply?

A.Set top_p to 0.1

B.Enable safety filters

C.Use few-shot prompting with examples of concise descriptions

D.Increase temperature to 0.9

AnswerC

Guides the model to match the style of provided examples.

Why this answer

The developer needs to constrain the model's output to be concise and relevant. Few-shot prompting provides the model with explicit examples of the desired output format (concise descriptions), guiding it to mimic that style and length. This directly addresses verbosity and irrelevant details without altering the model's fundamental randomness or safety settings.

Exam trap

The trap here is that candidates confuse hyperparameter tuning (top_p, temperature) with prompt engineering techniques, assuming that reducing randomness (top_p) or increasing creativity (temperature) can fix verbosity, when only explicit examples in the prompt can reliably enforce a specific output style.

How to eliminate wrong answers

Option A is wrong because setting top_p to 0.1 reduces the cumulative probability threshold for token sampling, which makes the output less diverse and more deterministic, but it does not teach the model to be concise or omit irrelevant details—it only narrows the pool of possible next tokens. Option B is wrong because safety filters block harmful or sensitive content (e.g., toxicity, violence), not verbose or irrelevant details; they do not control output length or relevance. Option D is wrong because increasing temperature to 0.9 increases randomness and creativity in token selection, which would likely make the output even more verbose and include more irrelevant details, the opposite of what is needed.

Full explanation →

645

Multi-Selectmedium

A developer wants to use Gemini for multimodal tasks involving images and text. Which two features are available via the Gemini API on Vertex AI? (Choose two.)

Select 2 answers

A.Fine-tuning with images

B.Image generation

C.Function calling

D.Code execution

E.Multimodal understanding (image+text)

AnswersC, E

Function calling is supported on Vertex AI.

Why this answer

Gemini on Vertex AI supports multimodal input (image+text) and function calling.

Full explanation →

646

Multi-Selecthard

Which THREE considerations are critical when deploying a generative AI model using Vertex AI Endpoints for a latency-sensitive application? (Choose THREE.)

Select 3 answers

A.Model size and architecture

B.Number of model versions

C.GPU type and number

D.Autoscaling configuration

E.Number of model instances

AnswersA, C, D

Larger models introduce higher latency.

Why this answer

Model size and architecture directly impact inference latency because larger models with more parameters require more computation per request. For latency-sensitive applications, choosing a smaller or distilled model (e.g., Gemma 2B vs. 27B) or using quantization can reduce response times. Vertex AI Endpoints serve the model as-is, so the model's inherent computational cost is the primary driver of per-request latency.

Exam trap

Google Cloud often tests the distinction between configuration choices that affect latency (GPU type, autoscaling, model size) versus operational or lifecycle management choices (version count, manual instance count) that do not directly impact per-request response time.

Full explanation →

647

MCQmedium

A company needs to deploy a chatbot on a mobile device that must work offline. They want to use Gemini for natural language understanding but need minimal latency and no cloud dependency. Which Gemini model variant is most appropriate?

A.Gemini Flash

B.Gemini Nano

C.Gemini Pro

D.Gemini Ultra

AnswerB

Gemini Nano is designed for on-device execution, enabling offline operation.

Why this answer

Gemini Nano is the most appropriate variant because it is specifically designed for on-device deployment, enabling offline operation with minimal latency. It is optimized for mobile devices through quantization and efficient architecture, allowing natural language understanding without any cloud dependency.

Exam trap

The trap here is that candidates often confuse 'lightweight' cloud models like Gemini Flash with truly on-device models like Gemini Nano, assuming that any 'fast' or 'small' model can work offline without understanding the fundamental requirement of local execution.

How to eliminate wrong answers

Option A is wrong because Gemini Flash is a lightweight cloud-based model optimized for speed and cost, but it still requires an internet connection to access Google's servers, making it unsuitable for offline use. Option C is wrong because Gemini Pro is a mid-tier cloud model designed for high-quality responses in cloud environments, not for on-device or offline scenarios. Option D is wrong because Gemini Ultra is the largest and most capable cloud model, intended for complex tasks with cloud infrastructure, and cannot run on a mobile device offline due to its massive computational requirements.

Full explanation →

648

MCQhard

A social media company is using a generative AI model to automatically moderate user-uploaded images for harmful content. They need to comply with the EU AI Act's requirements for high-risk AI systems. Which combination of actions is MOST appropriate?

A.Deploy the model without any filters, but log all decisions for audit

B.Use Google's safety filters, allow users to appeal automated decisions, and document the model's capabilities and limitations

C.Only rely on human moderators and disable AI moderation

D.Use a third-party model that has been certified by a notified body

AnswerB

Safety filters reduce harmful outputs, appeal mechanisms provide human oversight, and documentation satisfies transparency obligations under the EU AI Act.

Why this answer

The EU AI Act mandates risk management, transparency, and human oversight for high-risk systems. Google's AI Principles align with these requirements. The correct answer involves using safety filters for content, maintaining human review for appeals, and documenting the system's purpose and limitations.

Full explanation →

649

MCQhard

Refer to the exhibit. An administrator creates this IAM policy for a Vertex AI project. What is the effect of this policy?

A.Alice can view models; Bob can delete models

B.Alice can deploy pre-trained models; Bob can create and manage custom model code

C.Both have full access to all Vertex AI resources

D.Alice can train models; Bob can deploy models

AnswerB

aiplatform.user includes deployment permissions; customCodeModelAdmin covers custom code management.

Why this answer

The IAM policy grants Alice the `aiplatform.models.get` permission (allowing her to view and deploy pre-trained models) and grants Bob the `aiplatform.models.create` and `aiplatform.models.update` permissions (allowing him to create and manage custom model code). The policy uses separate bindings for each user, with specific roles that align with these actions.

Exam trap

Google Cloud often tests the distinction between specific IAM permissions (e.g., `get` vs. `create` vs. `delete`) and the common misconception that viewing a model implies full access or that creating a model implies the ability to deploy it.

How to eliminate wrong answers

Option A is wrong because Alice is granted `aiplatform.models.get`, which allows viewing models but not deleting them; Bob is granted `aiplatform.models.create` and `aiplatform.models.update`, which allow creating and updating models but not deleting them. Option C is wrong because the policy does not grant full access to all Vertex AI resources; it only grants specific permissions on models, and neither user has permissions for other resources like datasets or endpoints. Option D is wrong because Alice's permission (`aiplatform.models.get`) does not include training models, and Bob's permissions (`aiplatform.models.create` and `aiplatform.models.update`) are for custom model code management, not deploying models.

Full explanation →

650

MCQmedium

A machine learning engineer is evaluating a generative AI model for bias. They have a diverse test set covering gender, race, and age groups. Which metric would best indicate if the model's performance is systematically worse for certain demographic groups?

A.Model perplexity on held-out data

B.Equalized odds across demographic groups

C.Overall accuracy on the test set

D.Area under the ROC curve (AUC)

AnswerB

Equalized odds checks for fairness by comparing error rates across groups.

Why this answer

Equalized odds measures whether a model's predictions have equal false positive/negative rates across groups. The other options either measure different aspects or are not specific to fairness.

Full explanation →

651

MCQmedium

A company is using Vertex AI to generate email responses. They want to ensure sensitive customer data (PII) is not included in the output. What is the most effective approach?

A.Use a system prompt instructing the model to avoid PII.

B.Fine-tune the model on a dataset that excludes PII.

C.Manually review each output before sending.

D.Configure safety filters to block PII categories.

AnswerD

Safety filters can automatically block PII.

Why this answer

Safety filters in Vertex AI are specifically designed to block categories of harmful content, including PII, at the model's output layer. This provides a deterministic, automated guardrail that prevents sensitive data from being generated, unlike prompt-based instructions which can be overridden by the model's training. Safety filters operate on the model's response before it is returned, ensuring PII is caught even if the model attempts to generate it.

Exam trap

The trap here is that candidates assume a system prompt (Option A) is sufficient to control model behavior, but this exam tests the understanding that prompts are not enforceable guardrails, whereas safety filters are a hard technical control in Google Cloud.

How to eliminate wrong answers

Option A is wrong because system prompts are merely instructions and do not guarantee the model will comply; the model can still generate PII due to its training data or adversarial inputs. Option B is wrong because fine-tuning on a dataset that excludes PII does not prevent the model from generating PII from its pre-trained knowledge, and fine-tuning is costly and may not cover all edge cases. Option C is wrong because manual review is not scalable, introduces latency, and is prone to human error, making it ineffective for high-volume email generation.

Full explanation →

652

MCQhard

Refer to the exhibit. This JSON describes a Vertex AI endpoint with a deployed model. Which statement about scaling is true?

A.The endpoint uses only dedicated resources, no automatic scaling

B.The endpoint will automatically scale based on GPU utilization

C.The endpoint will scale from 1 to 3 replicas based on load using automatic scaling

D.The endpoint can scale to zero when not in use

AnswerA

DedicatedResources with min/max replicas means manual scaling.

Why this answer

The JSON shows that the endpoint is configured with `dedicatedResources` and no `autoscalingMetricSpecs` or `minReplicaCount`/`maxReplicaCount` fields. In Vertex AI, when you specify only `machineSpec` and a fixed `minReplicaCount` (here implicitly 1) without a `maxReplicaCount` or autoscaling metrics, the endpoint uses dedicated resources with no automatic scaling — the model will always run on exactly the number of replicas you define, regardless of load.

Exam trap

Google Cloud often tests the misconception that any endpoint with a `minReplicaCount` and `maxReplicaCount` automatically enables scaling, but the trap here is that without `autoscalingMetricSpecs`, the endpoint uses dedicated resources and does not scale dynamically — the `maxReplicaCount` is ignored if autoscaling metrics are absent.

How to eliminate wrong answers

Option B is wrong because Vertex AI automatic scaling is based on CPU utilization or custom metrics, not GPU utilization; GPU utilization is not a supported metric for autoscaling in Vertex AI endpoints. Option C is wrong because the JSON does not include `autoscalingMetricSpecs` or a `maxReplicaCount` field, which are required to enable automatic scaling from a minimum to a maximum number of replicas; without these, the endpoint uses a fixed replica count. Option D is wrong because Vertex AI endpoints with dedicated resources cannot scale to zero; scaling to zero is only possible with private endpoints using manual scaling or when using Vertex AI Prediction with a custom container that supports scale-to-zero, but dedicated resources always maintain at least one replica.

Full explanation →

653

MCQmedium

A company uses Vertex AI Agent Builder to create a customer support agent. They need the agent to answer questions about order status by calling an internal API. Which Vertex AI feature should they use?

A.Vertex AI RAG Engine

B.Vertex AI Extensions

C.Grounding with Google Search

D.Vertex AI Model Garden

AnswerB

Extensions enable the agent to call custom APIs as tools.

Why this answer

Extensions in Vertex AI Agent Builder allow the agent to call external APIs (including internal ones) as tools during conversation.

Full explanation →

654

MCQhard

An organization needs to deploy a generative AI application with strict compliance requirements, including data residency and auditability of model decisions. Which Google Cloud feature should they prioritize?

A.Colab Enterprise

B.Gemini API

C.Vertex AI

D.Model Garden

AnswerC

Vertex AI offers deployment options with data residency, audit logging, and governance features.

Why this answer

Vertex AI provides enterprise controls including data residency options, audit logs, and model monitoring for compliance. Other options are important but do not directly address data residency and auditability comprehensively.

Full explanation →

655

MCQeasy

A medical imaging team wants to generate synthetic X-ray images to augment a training dataset for a rare disease. Which type of generative model is most suitable for generating high-fidelity, realistic medical images?

A.Generative Adversarial Network (GAN)

B.Diffusion model

C.Variational Autoencoder (VAE)

D.Autoregressive transformer (e.g., PixelCNN)

AnswerB

Diffusion models currently produce the highest quality images.

Why this answer

Diffusion models are the most suitable for generating high-fidelity, realistic medical images because they iteratively denoise random noise into a coherent image through a learned reverse diffusion process, which produces superior sample quality and diversity compared to GANs, especially for complex, high-dimensional data like X-rays. Their training stability and ability to model fine-grained anatomical details without mode collapse make them the current state-of-the-art for medical image synthesis.

Exam trap

Google Cloud often tests the misconception that GANs are the default choice for image generation due to their popularity, but the trap here is that for high-fidelity medical imaging, diffusion models are preferred because they avoid GANs' mode collapse and training instability, which are critical in safety-sensitive domains.

How to eliminate wrong answers

Option A is wrong because GANs, while capable of generating realistic images, suffer from training instability, mode collapse, and difficulty in capturing the full diversity of medical image distributions, often producing artifacts that are unacceptable in clinical contexts. Option C is wrong because VAEs generate blurry and less detailed images due to their reliance on a variational lower bound and a Gaussian prior, which fails to capture the sharp edges and fine textures critical in X-ray images. Option D is wrong because autoregressive transformers like PixelCNN generate images pixel-by-pixel, which is computationally prohibitive for high-resolution medical images and lacks the global coherence and efficiency of diffusion models.

Full explanation →

656

MCQeasy

Which Google Cloud AI service would you use to transcribe customer service call recordings into text for subsequent analysis?

A.Speech-to-Text

B.Text-to-Speech

C.Translation API

D.Document AI

AnswerA

Speech-to-Text transcribes audio into text.

Why this answer

Speech-to-Text (STT) is the correct service because it is specifically designed to convert audio speech into written text using automatic speech recognition (ASR) models. For customer service call recordings, STT can handle domain-specific vocabulary, multiple speakers, and various audio formats, enabling downstream analysis like sentiment analysis or keyword extraction.

Exam trap

The trap here is confusing Speech-to-Text with Text-to-Speech or assuming that Translation API can handle audio input, when in fact it only works on text, leading candidates to pick a service that does not perform audio transcription.

How to eliminate wrong answers

Option B (Text-to-Speech) is wrong because it converts text into spoken audio, the reverse of what is needed for transcribing recordings. Option C (Translation API) is wrong because it translates text between languages but does not perform speech recognition or transcription from audio. Option D (Document AI) is wrong because it processes scanned documents and PDFs for text extraction and layout analysis, not audio files.

Full explanation →

657

MCQmedium

A financial services company uses a generative AI model to summarize customer complaints. They notice that summaries for certain demographics consistently omit negative sentiment. Which responsible AI practice should they apply FIRST to address this bias?

A.Store all prompts and responses in Cloud Logging for auditing

B.Implement SynthID watermarking on all generated summaries

C.Reduce the temperature parameter of the LLM to 0.1 to make outputs more deterministic

D.Evaluate the model's outputs for bias using a diverse test set that represents all customer demographics

AnswerD

Bias evaluation with representative data helps identify and quantify unfair bias, allowing the team to take corrective action.

Why this answer

Evaluating the model's outputs for bias using diverse test sets is essential to identify and mitigate unfair bias, as outlined in Google's AI Principles.

Full explanation →

658

MCQmedium

A company's generative AI model is producing biased outputs. What is the most effective mitigation strategy?

A.Use a larger model with more parameters to improve overall accuracy

B.Fine-tune the model using a balanced, representative dataset and implement output filtering

C.Use prompt engineering to instruct the model to avoid biased language

D.Increase the diversity of input samples by random sampling

AnswerB

Balanced data reduces bias during training, and filters catch biased outputs in production.

Why this answer

Fine-tuning on a balanced, representative dataset directly addresses the root cause of biased outputs by correcting the model's learned associations, while output filtering provides a safety net to catch residual bias. This combination is more effective than superficial fixes because it modifies the model's internal weights rather than just masking outputs.

Exam trap

Google Cloud often tests the misconception that prompt engineering or model scaling alone can fix bias, when in fact only retraining or fine-tuning with balanced data addresses the underlying weight distribution.

How to eliminate wrong answers

Option A is wrong because increasing model size does not inherently reduce bias; larger models can amplify biases present in training data due to higher capacity to memorize spurious correlations. Option C is wrong because prompt engineering only provides a surface-level instruction that the model may ignore or fail to generalize, especially if the bias is deeply embedded in its parameters. Option D is wrong because random sampling of inputs does not address the model's biased internal representations; it only diversifies the prompts, not the training data that caused the bias.

Full explanation →

659

MCQeasy

A data analyst wants to use Gemini in Google Sheets to help with complex formulas. Which feature should they use?

A.Model Garden in Vertex AI

B.Smart Compose in Gmail

C.Gemini for Workspace in Google Sheets

D.Help me write in Google Docs

AnswerC

Gemini in Sheets offers formula suggestions and assistance directly within the spreadsheet.

Why this answer

Gemini for Workspace in Google Sheets provides an AI-powered side panel that can generate, explain, and debug complex formulas directly within the spreadsheet environment. This feature is specifically designed to assist with formula creation and data analysis tasks, making it the correct choice for a data analyst using Gemini in Google Sheets.

Exam trap

The trap here is that candidates may confuse general-purpose AI writing features (like Help me write in Docs or Smart Compose in Gmail) with the specialized, context-aware formula assistance provided by Gemini for Workspace in Sheets, failing to recognize that each Workspace tool has a domain-specific integration.

How to eliminate wrong answers

Option A is wrong because Model Garden in Vertex AI is a repository of foundation models for building and deploying custom AI applications, not a feature integrated into Google Sheets for formula assistance. Option B is wrong because Smart Compose in Gmail is a feature for suggesting complete sentences in email composition, unrelated to spreadsheet formulas or data analysis. Option D is wrong because Help me write in Google Docs is a generative writing assistant for document creation, not designed to handle complex formulas or spreadsheet-specific tasks.

Full explanation →

660

MCQmedium

You are a generative AI architect for a large e-commerce company. Your team has built a product description generator using Vertex AI's text-bison model. The model is accessed via the Vertex AI API from a web application. You have set the temperature to 0.5 and top_k to 40. The team reports that the generated descriptions are often too generic and lack creativity. They want the descriptions to be more diverse and engaging. You are also concerned about cost, as each API call is billed. Which change should you recommend to increase creativity while managing cost?

A.Keep temperature at 0.5 but reduce top_k to 20.

B.Increase the temperature to 0.8 and keep top_k at 40.

C.Switch to a larger model like text-bison@002 and keep same parameters.

D.Decrease the temperature to 0.2 and increase top_k to 60.

AnswerB

Higher temperature increases diversity and creativity.

Why this answer

Increasing the temperature to 0.8 makes the model's output probability distribution flatter, which increases randomness and allows less likely tokens to be selected. This directly addresses the need for more diverse and creative descriptions. Keeping top_k at 40 ensures the model still considers a broad set of candidate tokens, balancing creativity with coherence, and does not increase API call costs since temperature and top_k are inference parameters that do not affect billing.

Exam trap

Google Cloud often tests the misconception that increasing creativity requires a larger model or more expensive resources, when in fact tuning sampling parameters like temperature and top_k is the correct, cost-neutral approach.

How to eliminate wrong answers

Option A is wrong because reducing top_k to 20 narrows the set of candidate tokens, which actually reduces diversity and can make outputs more generic, counteracting the goal of increasing creativity. Option C is wrong because switching to a larger model like text-bison@002 would increase cost per API call (larger models are billed at higher rates) without guaranteeing more creativity; creativity is controlled by sampling parameters, not model size alone. Option D is wrong because decreasing temperature to 0.2 makes the model more deterministic and conservative, reducing creativity, and increasing top_k to 60 does not compensate for the loss of randomness — the net effect is less diverse outputs.

Full explanation →

661

MCQeasy

A product manager wants to add a feature that drafts meeting summaries automatically in Google Meet. Which Gemini for Google Workspace capability should they use?

A.Vertex AI Model Garden

B.Gemini for Workspace in Google Meet

C.Vertex AI Agent Builder

D.Duet AI in Google Slides

AnswerB

Gemini for Workspace (formerly Duet AI) includes meeting summaries in Google Meet.

Why this answer

Gemini for Workspace in Google Meet provides meeting summaries. Duet AI for Slides does not generate meeting summaries. Vertex AI Agent Builder and Model Garden are not directly integrated into Meet.

Full explanation →

662

MCQeasy

A developer is using the Gemini API to build a chatbot. They want the model to always respond in a friendly, professional tone. Which prompt engineering technique should they use?

A.Set system instructions to 'You are a friendly and professional assistant.'

B.Include a few-shot example in every user message.

C.Set the temperature to 0.2.

D.Set max output tokens to 100.

AnswerA

System instructions define the assistant's behavior for the entire session.

Why this answer

Setting system instructions is the most direct and reliable way to define the model's persona and behavioral constraints. In the Gemini API, system instructions act as a persistent, top-level directive that influences every response, ensuring the chatbot consistently adopts a friendly and professional tone without requiring repeated examples or parameter tuning.

Exam trap

Google Cloud often tests the distinction between controlling output style (system instructions) versus controlling output randomness (temperature) or length (max tokens), so the trap here is that candidates may confuse temperature or token limits with persona control, thinking that lowering creativity or capping length will enforce a specific tone.

How to eliminate wrong answers

Option B is wrong because including a few-shot example in every user message is inefficient and not a persistent technique; it would require repeating the example in each turn, increasing token usage and latency, and it does not guarantee consistent tone across all interactions. Option C is wrong because setting the temperature to 0.2 controls randomness and creativity, not tone; a low temperature makes outputs more deterministic but does not enforce a specific persona or style. Option D is wrong because setting max output tokens to 100 limits response length but has no effect on the tone or style of the output; it only truncates the response.

Full explanation →

663

Multi-Selectmedium

A company wants to integrate generative AI into their existing CRM workflow to draft personalized email responses. They have limited engineering resources. Which two approaches should they consider? (Choose TWO)

Select 2 answers

A.Use Vertex AI API with a low-code integration platform (e.g., Apigee)

B.Fine-tune a model on historical email data to ensure brand voice

C.Use Gemini API via Google Apps Script to add a custom menu in the CRM

D.Deploy a dedicated GPU cluster for inference

E.Build a custom web UI for the assistant from scratch

AnswersA, C

Low-code platforms reduce the need for custom coding.

Why this answer

Using Gemini API via Apps Script is a lightweight integration, and using Vertex AI API with a low-code tool like Apigee or Cloud Functions can also minimize engineering effort. Building a custom UI or fine-tuning is resource-intensive.

Full explanation →

664

Multi-Selecthard

A manufacturing company wants to use GenAI to generate maintenance reports from sensor data. They need structured output (JSON) for downstream systems, and they want to reduce token costs. Which THREE strategies should they use?

Select 3 answers

A.Use batch API requests for multiple sensor readings

B.Use the largest available model to ensure accuracy

C.Use structured output formatting in the prompt (e.g., 'Return JSON')

D.Choose the smallest model that meets accuracy requirements

E.Include multiple few-shot examples of JSON in every prompt

AnswersA, C, D

Batch requests reduce per-token cost.

Why this answer

Structured output ensures JSON format; batch requests reduce cost; the smallest suitable model minimizes token usage. Few-shot adds tokens; caching may not help for diverse sensor data.

Full explanation →

665

Multi-Selectmedium

A company is designing a prompt engineering strategy for a customer service chatbot using Gemini. Which two practices are recommended for improving response quality? (Choose TWO)

Select 2 answers

A.Use chain-of-thought prompting

B.Always provide multiple examples in the prompt

C.Avoid any context in the prompt

D.Set temperature to 1.0 for maximum creativity

E.Include a system instruction to define the role

AnswersA, E

Chain-of-thought encourages logical reasoning, improving accuracy.

Why this answer

Chain-of-thought prompting (A) is recommended because it guides the model to reason step-by-step, improving accuracy on complex customer service queries by breaking down multi-step problems. This technique leverages Gemini's ability to follow logical sequences, reducing errors in tasks like troubleshooting or escalation decisions.

Exam trap

Google Cloud often tests the misconception that higher temperature always improves creativity, but in customer service, lower temperature is critical for deterministic, safe responses, and candidates may overlook the role of system instructions in defining behavior.

Full explanation →

666

MCQmedium

A company wants to integrate GenAI into their existing customer relationship management (CRM) system. The CRM is a third-party SaaS application. Which implementation pattern is MOST suitable?

A.API-first integration by calling Vertex AI API from the CRM's custom code

B.Building a standalone AI application and exporting data manually

C.Embedding GenAI using Google Workspace add-ons

D.Using Apps Script to extend Google Sheets connected to the CRM

AnswerA

The CRM can make API calls to Vertex AI for predictions, keeping the CRM intact.

Why this answer

API-first integration allows the CRM to call Google Cloud GenAI APIs without modifying the CRM's core. Workspace add-ons are for Google Workspace, not SaaS CRM.

Full explanation →

667

MCQeasy

A marketing agency wants to generate images using Imagen on Vertex AI. They need to ensure the images are unique and avoid copyright issues. Which parameter adjustment is most relevant?

A.Increase training steps

B.Increase seed variability

C.Use negative prompts

D.Set safety threshold

AnswerC

Specifies elements to avoid, reducing copyright risk.

Why this answer

Negative prompts allow the model to exclude specific concepts, styles, or elements from generated images, directly reducing the risk of replicating copyrighted or trademarked content. By explicitly telling Imagen what not to include, the agency can steer outputs away from protected works without needing to modify training data or safety filters.

Exam trap

Google Cloud often tests the distinction between safety filters (which block harmful content) and negative prompts (which control stylistic or conceptual exclusion), leading candidates to mistakenly choose safety threshold adjustments for copyright avoidance.

How to eliminate wrong answers

Option A is wrong because increasing training steps does not affect the uniqueness or copyright compliance of outputs; it only refines model convergence on the existing training distribution. Option B is wrong because seed variability controls randomness in latent noise initialization, not the semantic content of the image, so it cannot prevent copyright infringement. Option D is wrong because safety thresholds filter harmful or policy-violating content (e.g., violence, hate speech), not copyrighted or trademarked elements.

Full explanation →

668

MCQeasy

A developer needs to integrate a GenAI model into an existing customer relationship management (CRM) system. The CRM exposes REST APIs and runs on-premises. Which integration pattern is MOST suitable?

A.Use Apps Script to call the model from Google Sheets

B.Fine-tune a model on CRM data and deploy it on-premises

C.Implement an API-first integration by calling Vertex AI API from the CRM's backend

D.Build a custom Google Workspace add-on

AnswerC

Vertex AI API can be called via HTTPS from any system that supports REST, making it ideal for integrating with an on-premises CRM.

Why this answer

API-first integration using Vertex AI API allows any system with HTTP capabilities to call the GenAI model. Workspace add-ons and Apps Script are for Google Workspace only. Fine-tuning doesn't help with integration.

Full explanation →

669

MCQhard

A financial services firm wants to use Gemini to analyze customer support transcripts and generate summaries. Compliance requires that the model never output any personally identifiable information (PII). Which combination of techniques should they implement?

A.Configure Gemini safety settings to block PII and use a separate PII detection API for post-processing

B.Fine-tune Gemini on redacted transcripts and rely on the model to not generate PII

C.Use a smaller model that has never seen PII in training

D.Only use prompt instructions telling the model to avoid PII

AnswerA

Safety filters reduce PII in outputs, and a post-processing API (e.g., DLP) redacts any remaining PII.

Why this answer

Using Gemini with safety filters and a post-processing step to redact PII provides defense in depth. Fine-tuning on redacted data might not cover all cases, and prompt instructions alone are not reliable.

Full explanation →

670

MCQhard

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?

A.Increase the batch size for batch requests

B.Enable context caching for repeated prompt prefixes

C.Switch to a smaller model size

D.Ignore the cost increase as it will stabilize

AnswerB

Context caching stores repeated prefixes so they are not re-sent with each request, reducing token consumption and cost.

Why this answer

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.

Full explanation →

671

MCQhard

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?

A.Cloud Storage

B.Persistent Disk

C.Cloud Filestore

D.Cloud Spanner

AnswerA

Cloud Storage provides high throughput for large datasets, especially with parallel reads.

Why this answer

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.

Exam trap

Google Cloud often tests the misconception that local or attached block storage (Persistent Disk) is faster than object storage for ML training, but candidates fail to recognize that TPU training requires distributed, parallel data access that object storage (Cloud Storage) uniquely provides at scale.

How to eliminate wrong answers

Option B is wrong because Persistent Disk is a block storage device with a maximum throughput of ~1.2 GB/s per instance (for pd-ssd), which is far below the multi-GB/s requirements of TPU training and cannot scale horizontally across many workers without complex striping. Option C is wrong because Cloud Filestore is a managed NFS filestore that introduces network latency and has throughput caps (e.g., 1.2 GB/s for the Basic tier, 4.8 GB/s for the High Scale tier), making it unsuitable for the high-throughput, low-latency data streaming needed by TPUs. Option D is wrong because Cloud Spanner is a globally distributed relational database service designed for transactional consistency and ACID compliance, not for high-throughput sequential read of training data; its throughput is limited by node count and query overhead, and it is not a file storage solution.

Full explanation →

672

MCQhard

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?

A.Use Document AI to convert PDFs to structured text, then use Vertex AI with a prompt that specifies clauses to extract

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

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

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

AnswerA

Document AI handles layout parsing, and the structured text is then fed to a GenAI model with a well-designed prompt. This combination is scalable and cost-effective.

Why this answer

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.

Full explanation →

673

Multi-Selecthard

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

Select 3 answers

A.Using only English prompts

B.Increasing model size

C.Using a more diverse training dataset

D.Using safety filters

E.Applying prompt engineering to instruct the model to be fair

AnswersC, D, E

Diverse data reduces the risk of model learning biased patterns.

Why this answer

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.

Exam trap

Google Cloud often tests the misconception that increasing model size or using a single language (like English) can solve bias, when in reality these actions can worsen bias by amplifying existing skews or introducing new cultural blind spots.

Full explanation →

674

MCQmedium

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?

A.Using prompt engineering with in-context learning

B.Full fine-tuning of all model parameters

C.Training a new model from scratch on the domain data

D.Adapter-based fine-tuning using LoRA

AnswerD

LoRA is parameter-efficient, cost-effective, and reduces forgetting.

Why this answer

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.

Exam trap

The Generative AI Leader exam often tests the misconception that prompt engineering (Option A) is a form of fine-tuning, when in fact it is a zero-shot or few-shot inference technique that does not modify model parameters, making it unsuitable for persistent domain adaptation with limited labeled data.

How to eliminate wrong answers

Option A is wrong because prompt engineering with in-context learning does not update model weights, so it cannot adapt the model to a specialized domain with limited labeled data in a persistent manner; it relies on the model's existing knowledge and context window, which is insufficient for deep domain adaptation. Option B is wrong because full fine-tuning of all model parameters updates the entire model, which is computationally expensive and, with limited labeled data, risks catastrophic forgetting of the original pre-trained knowledge. Option C is wrong because training a new model from scratch on domain data requires a massive amount of labeled data and compute resources, defeating the purpose of leveraging a pre-trained LLM and reducing cost.

Full explanation →

675

MCQhard

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?

A.Fine-tune the model with RLHF

B.Set the temperature to 0.0

C.Implement retrieval-augmented generation (RAG) with a curated knowledge base

D.Use prompt engineering to instruct the model to be accurate

AnswerC

RAG grounds outputs in retrieved facts, reducing hallucinations.

Why this answer

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.

Exam trap

The trap here is that candidates often choose 'Set the temperature to 0.0' because they confuse reducing randomness with eliminating factual errors, but temperature only controls output variability, not the truthfulness of the model's internal knowledge.

How to eliminate wrong answers

Option A is wrong because RLHF (Reinforcement Learning from Human Feedback) optimizes the model for human preference alignment and helpfulness, but it does not provide a mechanism to retrieve or verify facts from an external source, so it cannot reliably prevent hallucinations in factual domains. Option B is wrong because setting temperature to 0.0 makes the model deterministic (always picking the highest-probability token), but it does not correct factual errors stored in the model's weights; the model can still confidently generate false information. Option D is wrong because prompt engineering instructs the model to be accurate, but it is a soft constraint that the model can easily override; without external grounding, the model has no way to verify its own output against a trusted source.

Full explanation →

Page 9 of 10

All pages

1 2 3 4 5 6 7 8 9 10

Practice Generative AI Leader by domain

Target a specific domain to shore up weak areas.

Fundamentals of Generative AI Business Strategies for Generative AI Solutions Google Cloud's Generative AI Offerings Techniques to Improve Generative AI Model Output

See all domains with question counts →