683 questions total · 10pages · All types, answers revealed
A startup wants to build a text-to-speech application for generating audiobooks. Which Google Cloud generative AI service is best suited for this task?
Chirp is Google's model for generating natural-sounding speech from text.
Why this answer
Chirp is Google Cloud's text-to-speech AI model. Imagen generates images, Codey generates code, and Gemini is a multimodal model not specialized for high-quality speech synthesis.
A company wants to build a chatbot that can answer questions about recent news events up to the current date. They want to minimize hallucinations and ensure factual accuracy. Which Google Cloud feature should they integrate?
Search grounding provides up-to-date information from Google Search.
Why this answer
The Gemini API with Google Search grounding enables the chatbot to retrieve and cite real-time information from the web, directly addressing the need for up-to-date news and factual accuracy. Grounding connects the model's responses to verified search results, reducing hallucinations by anchoring answers in current, authoritative sources rather than relying solely on the model's training data.
Exam trap
The Generative AI Leader exam often tests the distinction between static NLP capabilities (entity analysis, document parsing) and dynamic retrieval-augmented generation, leading candidates to choose a familiar but incorrect service like Natural Language AI or Document AI instead of recognizing that grounding requires a search-backed integration.
How to eliminate wrong answers
Option B is wrong because Natural Language AI entity analysis extracts entities (e.g., names, dates) from text but does not provide real-time grounding or fact-checking against current events; it is a static analysis tool, not a retrieval mechanism. Option C is wrong because Document AI is designed for parsing and extracting data from documents (e.g., invoices, forms), not for answering questions about live news or integrating with web search. Option D is wrong because Vertex AI Vector Search enables semantic similarity search over a fixed vector index of pre-ingested data; it cannot access or verify real-time news events unless the index is continuously updated, and it lacks the grounding mechanism to minimize hallucinations from stale or incomplete data.
A gen AI application produces hallucinations (factually incorrect outputs). Which mitigation strategy is LEAST effective?
Higher temperature leads to more diverse but less predictable outputs, exacerbating hallucinations.
Why this answer
Increasing model temperature makes the model more random and creative, which directly increases the likelihood of hallucinations. It does not constrain or ground the output in factual data, making it the least effective mitigation strategy among the options.
Exam trap
Google Cloud often tests the misconception that increasing model temperature improves accuracy by making the model 'more confident,' when in reality it increases randomness and hallucination risk.
How to eliminate wrong answers
Option A is wrong because prompt templates with constraints (e.g., specifying 'only answer from the provided context') reduce the model's freedom to generate unverified content, thereby lowering hallucination risk. Option B is wrong because grounding with a knowledge base ties the model's outputs to verified facts, preventing fabrication by restricting the response to a trusted data source. Option C is wrong because retrieval-augmented generation (RAG) explicitly fetches relevant documents from a knowledge base before generation, ensuring the output is based on retrieved evidence rather than parametric memory alone.
A team is evaluating generative AI models on Vertex AI. They need to compare models based on specific criteria. Which TWO criteria are most important for selecting a model for a text summarization task?
ROUGE evaluates summary quality against references.
Why this answer
ROUGE scores are the standard evaluation metric for text summarization tasks, measuring the overlap of n-grams, word sequences, and word pairs between generated summaries and reference summaries. This directly quantifies summary quality, making it the most important criterion for model selection.
Exam trap
Google Cloud often tests the misconception that model size or cost are primary selection criteria, when in fact task-specific metrics like ROUGE are the correct focus for evaluating generative model output quality.
A machine learning engineer is deploying a large generative model on Vertex AI. The model requires a GPU with high memory. Which machine configuration should they choose?
A2 machines offer A100s with large memory, suitable for large models.
Why this answer
The a2-highgpu-4g machine series is specifically designed for large-scale GPU-accelerated workloads, offering 4 NVIDIA A100 GPUs with 40GB of high-bandwidth memory (HBM2e) each, totaling 160GB of GPU memory. This configuration provides the high memory capacity required for training or serving large generative models, such as LLMs or diffusion models, which often exceed the memory limits of smaller GPUs.
Exam trap
The trap here is that candidates may choose a cheaper or single-GPU option (like C or D) without calculating the total GPU memory needed, or mistakenly think a CPU-only instance (A) can handle GPU-accelerated workloads, ignoring that large generative models require both high GPU memory and parallel processing.
How to eliminate wrong answers
Option A is wrong because c2-standard-16 is a compute-optimized machine without any GPU, which cannot provide the GPU memory needed for large generative models. Option C is wrong because n1-standard-4 with a single T4 GPU offers only 16GB of GPU memory, insufficient for large models that require tens or hundreds of gigabytes. Option D is wrong because n2-standard-8 with a single P4 GPU provides only 8GB of GPU memory, far below the requirements for large generative models and lacks the parallelism of multiple GPUs.
A company is using Vertex AI Gemini API to analyze customer feedback. They notice that the model occasionally generates offensive content. They have already set safety settings to block high-probability harmful content. What additional step should they take to further reduce offensive outputs?
Stricter thresholds block more offensive outputs.
Why this answer
The company has already blocked high-probability harmful content, but offensive outputs can still occur at lower probability thresholds. By adjusting safety settings to block medium-probability harmful content, they tighten the filter to catch more borderline cases without requiring model retraining or sacrificing output diversity. This leverages Vertex AI's configurable safety filters, which operate on likelihood categories (e.g., high, medium, low) rather than just binary blocking.
Exam trap
The trap here is that candidates assume fine-tuning (Option D) is the default fix for any output quality issue, but safety filtering is a separate, configurable layer that should be tuned before retraining, and temperature (Option A) is often mistakenly thought to control safety when it only controls randomness.
How to eliminate wrong answers
Option A is wrong because setting temperature to 0.0 makes the model deterministic and reduces creativity, but it does not filter or block offensive content; temperature controls randomness in token selection, not safety. Option C is wrong because context caching improves latency and cost for repeated prompts by storing context, but it has no effect on content safety or filtering harmful outputs. Option D is wrong because fine-tuning on customer feedback data could inadvertently reinforce biases or offensive patterns in the data, and it does not directly address safety filtering; safety settings are a separate, configurable layer that should be adjusted first.
Refer to the exhibit. A machine learning engineer is configuring a model using this YAML. What is the purpose of the 'tuningPipeline' field?
The tuningPipeline references a pipeline that performs supervised fine-tuning of the base model.
Why this answer
The 'tuningPipeline' field in this YAML configuration specifies a dedicated pipeline for fine-tuning the base model, which is a common practice in MLOps frameworks like Vertex AI Pipelines or Kubeflow. It allows the engineer to define a separate workflow for parameter-efficient fine-tuning (e.g., LoRA) or full fine-tuning, distinct from training from scratch or serving. This field is essential for orchestrating the fine-tuning process, including data preprocessing, training, and evaluation steps, without affecting the base model's original weights.
Exam trap
Google Cloud often tests the distinction between 'tuningPipeline' (fine-tuning an existing model) and 'trainingPipeline' (training from scratch), and the trap here is that candidates confuse fine-tuning with full training or assume the field is for inference tasks like prediction.
How to eliminate wrong answers
Option B is wrong because 'tuningPipeline' is not used for online prediction; online prediction is typically configured via a separate 'predict' or 'serving' pipeline or endpoint specification. Option C is wrong because 'tuningPipeline' is specifically for fine-tuning an existing base model, not for training from scratch, which would require a different pipeline definition (e.g., 'trainingPipeline') with full hyperparameter search. Option D is wrong because batch prediction is handled by a distinct 'batchPrediction' or 'batch' pipeline configuration, not by the tuning pipeline, which focuses on model adaptation rather than inference.
A healthcare organization wants to use generative AI for clinical decision support. They must comply with regulatory requirements for high-risk AI systems. Which THREE measures should they implement? (Choose 3)
High-risk AI systems require human oversight to ensure safety and accountability.
Why this answer
Human oversight, risk management, and transparency are key regulatory requirements for high-risk AI systems.
A media company uses Vertex AI to generate video captions. The generated captions sometimes contain factual errors about named entities (e.g., actor names). Which technique would most likely reduce these errors?
Grounding supplies factual context to the model.
Why this answer
Vertex AI grounding connects the model to a knowledge base of verified entities, allowing it to retrieve authoritative facts during generation. This reduces hallucinations about named entities by constraining outputs to validated data rather than relying solely on the model's parametric knowledge.
Exam trap
The trap here is that candidates confuse techniques that control output randomness (temperature, top_p) with techniques that improve factual accuracy, overlooking the fundamental need for external knowledge retrieval via grounding.
How to eliminate wrong answers
Option A is wrong because response caching stores previous outputs for reuse, which does not correct factual errors—it may even propagate them. Option B is wrong because increasing the temperature parameter increases randomness in token selection, making factual errors more likely, not less. Option D is wrong because decreasing top_p to 0.3 narrows the sampling pool to only the most probable tokens, which can reduce creativity but does not address factual accuracy about named entities—it still relies on the model's internal knowledge, which may be incorrect.
A sales team wants to use Gemini for Google Workspace to draft personalized emails to prospects. Which feature in Gmail should they use?
Smart Compose provides inline suggestions as you type.
Why this answer
Smart Compose in Gmail suggests complete sentences to help draft emails faster. 'Help me write' is in Docs, not Gmail.
A financial services firm is deploying a Gemini-based application that must comply with GDPR. The application processes customer queries and may include personal data. Which Google Cloud capability should the firm use to ensure that the model does not expose personally identifiable information (PII) in its responses?
DLP can be used to scan for PII and redact it before sending to the model or after receiving the response, ensuring GDPR compliance.
Why this answer
Sensitive Data Protection (formerly DLP) can be integrated into the GenAI pipeline to inspect and redact PII from prompts and responses, ensuring compliance.
A research team is fine-tuning a large language model on a dataset containing personal data of EU citizens. They must comply with GDPR. Which measure is ESSENTIAL?
Consent is a fundamental requirement under GDPR for processing personal data.
Why this answer
Under GDPR, processing personal data requires a lawful basis; explicit consent is essential when no other basis (e.g., legitimate interest) clearly applies, especially for fine-tuning where data is used to train a model that may memorize and regenerate personal information. Without consent, the processing is unlawful, exposing the team to significant fines and regulatory action. This is a foundational requirement that overrides technical measures like storage location or documentation.
Exam trap
The Generative AI Leader exam often tests the distinction between legally required measures (like consent) versus best-practice technical controls (like Model Cards or safety filters), leading candidates to pick a technically sound but legally insufficient option.
How to eliminate wrong answers
Option A is wrong because GDPR does not mandate a specific geographic region for storage; it allows data transfers to third countries under adequacy decisions or appropriate safeguards (e.g., Standard Contractual Clauses), so storage location alone is not essential. Option C is wrong because a Model Card is a best practice for transparency and documentation but is not a legal requirement under GDPR; it does not provide a lawful basis for processing personal data. Option D is wrong because safety filters address output harms (e.g., toxicity, bias) but do not address the legality of using personal data for training; they are a post-hoc mitigation, not a prerequisite for lawful processing.
Refer to the exhibit. A user with this IAM role tries to deploy a model to a Vertex AI Endpoint but fails. What is the most likely reason?
Deploying a model requires the aiplatform.deployer role or equivalent permissions.
Why this answer
The user has an IAM role but lacks the specific permission `aiplatform.deployments.create` required to deploy a model to a Vertex AI Endpoint. The `roles/aiplatform.deployer` role includes this permission, while the user's existing role does not, causing the deployment to fail. Even if the user can use other Vertex AI services, deploying a model to an endpoint is a distinct action that requires this specific role.
Exam trap
Google Cloud often tests the distinction between broad roles like `roles/aiplatform.admin` and specific roles like `roles/aiplatform.deployer`, trapping candidates who assume that any Vertex AI role can perform all actions, when in fact deployment requires a dedicated permission set.
How to eliminate wrong answers
Option A is wrong because the user is able to interact with Vertex AI (they have an IAM role), but the failure is specific to the deploy action, not a blanket denial of all Vertex AI access. Option B is wrong because model artifacts can be deployed to endpoints in any region as long as the endpoint exists; Vertex AI supports cross-region deployment by copying the model artifact to the endpoint's region automatically. Option D is wrong because the `roles/aiplatform.admin` role is overly permissive and includes full administrative access, which is not required for deploying a model; the principle of least privilege dictates that the `roles/aiplatform.deployer` role is sufficient and more appropriate.
A financial institution wants to deploy a gen AI model for fraud detection but must comply with strict regulations regarding explainability. What is the best strategy?
Interpretable models satisfy explainability requirements while maintaining reasonable performance.
Why this answer
Regulatory compliance for fraud detection demands explainability, which complex black-box models cannot provide. A smaller interpretable model (e.g., logistic regression or decision tree) offers transparency into decision factors, satisfying regulations like GDPR's right to explanation while maintaining acceptable accuracy for the use case.
Exam trap
Google Cloud often tests the misconception that post-hoc explainability tools (like Vertex AI Explainable AI) are equivalent to inherent model interpretability, leading candidates to choose complex models with added explanation layers instead of simpler, transparent models.
How to eliminate wrong answers
Option A is wrong because Vertex AI Explainable AI provides post-hoc explanations for complex models, but these approximations may not meet strict regulatory standards for full transparency and can be unreliable. Option B is wrong because ensembling multiple models increases complexity and opacity, making it harder to explain individual predictions and often violating explainability requirements. Option C is wrong because relying on external auditing for a large black-box model does not guarantee inherent explainability; auditors still face the same opacity, and regulations typically require model-inherent interpretability, not just external review.
A company deployed a Gemini model on Vertex AI for real-time inference. After a week, they notice that some requests return 500 Internal Server Error, and the endpoint is occasionally unreachable. The endpoint is configured with minReplicaCount=1 and maxReplicaCount=2. What is the most likely cause?
C is correct because the single replica cannot handle bursts, leading to errors.
Why this answer
With minReplicaCount=1 and maxReplicaCount=2, the endpoint starts with a single replica. Under peak load, that single replica can become saturated (CPU/memory exhaustion), causing 500 errors and unreachability. Autoscaling can add a second replica, but if the traffic spike is sudden or the scaling metric takes time to trigger, the single replica is overwhelmed before the second instance is provisioned.
Exam trap
The trap here is that candidates assume autoscaling instantly handles spikes, but they overlook the provisioning delay and the fact that a single replica can be overwhelmed before the second replica is ready.
How to eliminate wrong answers
Option A is wrong because autoscaling is not disabled; the minReplicaCount=1 and maxReplicaCount=2 configuration explicitly enables autoscaling between 1 and 2 replicas. Option B is wrong because updating a model while the endpoint is serving requests does not cause 500 errors or unreachability; Vertex AI supports model updates with zero-downtime via canary deployments or traffic splitting. Option D is wrong because reaching the Vertex AI endpoint quota would result in a 429 (Too Many Requests) or a quota-exceeded error, not a 500 Internal Server Error or intermittent unreachability.
Which TWO methods are most effective for improving factual accuracy in a language model's responses? (Choose two.)
Prompt engineering can explicitly direct the model to verify claims or stick to given knowledge.
Why this answer
Prompt engineering can explicitly instruct the model to base its responses on provided facts, reducing reliance on parametric knowledge that may be outdated or incorrect. By including directives like 'Use only the information in the following text' or 'Answer based solely on the provided context,' the model is guided to prioritize given facts over its internal training data, which improves factual accuracy in the output.
Exam trap
A common misconception is that reducing randomness (temperature) or increasing token diversity (top-k) directly improves factual accuracy, when in fact these parameters affect output style and creativity, not the correctness of the underlying facts.
A company is moving a GenAI proof-of-concept to production. The pilot showed promising results, but the production environment must handle higher traffic, ensure low latency, and manage costs. Which THREE actions should they take?
Caching reduces latency and cost for frequent requests.
Why this answer
A/B testing helps compare versions. Choosing a smaller model balances cost and latency. Caching reduces repeated work.
Fine-tuning is expensive and not necessary. A single endpoint without testing is risky.
A graphic design company wants to generate high-quality synthetic images for product mockups. Which Google Cloud generative AI service is most suitable?
Imagen is specifically built for image generation and is accessible via Vertex AI.
Why this answer
Imagen on Vertex AI is the correct choice because it is Google Cloud's state-of-the-art text-to-image diffusion model specifically designed to generate high-quality, photorealistic synthetic images from natural language prompts. This directly meets the requirement for creating product mockups, as Imagen can produce custom visuals with fine-grained control over style and composition, and it integrates seamlessly with Vertex AI for deployment and management.
Exam trap
The trap here is that candidates may confuse AutoML Vision's ability to classify or detect objects in images with generative image creation, leading them to select Option A despite it lacking any generative capability.
How to eliminate wrong answers
Option A is wrong because AutoML Vision is a traditional machine learning service for training custom image classification, object detection, or segmentation models on labeled datasets; it does not generate synthetic images from text prompts. Option C is wrong because Codey APIs are specialized for generating code snippets, documentation, and code completions, not for creating visual content like images. Option D is wrong because Natural Language API is designed for analyzing and extracting insights from text (e.g., sentiment, entity recognition), not for generating synthetic images.
A machine learning engineer needs to deploy a Gemini model on an Android device for offline inference (no internet connection) to provide real-time suggestions. Which Gemini model variant is MOST appropriate?
Nano is designed for on-device inference, small model size, suitable for offline usage.
Why this answer
Gemini Nano is specifically designed for on-device inference with small memory footprint. Ultra, Pro, and Flash are cloud-only models that require internet connectivity.
A project manager wants to measure the success of a generative AI feature that summarizes meeting transcripts. Which TWO metrics are MOST appropriate for evaluating quality improvement?
Accuracy directly measures output quality.
Why this answer
Accuracy of summaries and user satisfaction directly measure quality. AHT measures productivity. Cost per summary is financial.
API latency is performance.
A healthcare company wants to use generative AI to summarize patient records. They are concerned about data privacy and HIPAA compliance. Which Google Cloud feature should they use to protect patient data?
CMEK ensures data is encrypted with keys controlled by the customer, and VPC-SC prevents data exfiltration.
Why this answer
D is correct because Customer-managed encryption keys (CMEK) with VPC Service Controls provide a defense-in-depth approach for HIPAA compliance. CMEK allows the healthcare company to control and manage the encryption keys used to protect patient data at rest, while VPC Service Controls prevent data exfiltration by restricting data movement outside a defined service perimeter. This combination ensures that even if an attacker gains access, they cannot decrypt the data or move it out of the controlled environment, directly addressing data privacy and HIPAA requirements.
Exam trap
The trap here is that candidates often confuse data discovery and de-identification tools (DLP) with data protection and access control mechanisms (CMEK + VPC Service Controls), leading them to pick Cloud DLP API despite it not providing encryption or perimeter controls required for HIPAA compliance.
How to eliminate wrong answers
Option A is wrong because Cloud Audit Logs only record who did what, when, and where, but do not protect or encrypt patient data; they are a monitoring tool, not a data protection mechanism. Option B is wrong because Confidential VMs encrypt data in use using AMD SEV, but they do not control data exfiltration or provide the perimeter-based access controls needed for HIPAA compliance; they focus on memory encryption, not data movement restrictions. Option C is wrong because Cloud Data Loss Prevention (DLP) API is used for inspecting, classifying, and de-identifying sensitive data, but it does not provide encryption key management or network-level controls to prevent unauthorized data access or exfiltration.
A company is considering monetizing a generative AI-powered product. Which two business models are most common and viable?
Pay-per-use matches costs to usage, common in cloud API services.
Why this answer
Pay-per-use based on tokens consumed aligns directly with the operational cost structure of generative AI models, where each inference incurs compute and memory costs proportional to the number of tokens processed. This model allows customers to pay only for what they use, making it viable for variable workloads and avoiding upfront commitment, while providers can scale revenue with usage. It is the most common monetization strategy for API-based generative AI services, such as OpenAI's GPT-4 or Anthropic's Claude, where pricing is explicitly tied to token counts.
Exam trap
Google Cloud often tests the misconception that one-time licensing (Option B) is viable for AI models, but candidates must recognize that generative AI models are not static software—they require ongoing compute, updates, and scaling, making subscription or pay-per-use models the only sustainable approaches.
A data science team is fine-tuning a large language model using Vertex AI to generate marketing copy. They notice that the generated text is often repetitive and lacks creativity. Which technique should they apply to improve output diversity?
Higher temperature increases randomness and diversity in generated text.
Why this answer
Increasing the temperature parameter to 0.9 raises the randomness of the probability distribution over tokens, allowing less likely tokens to be selected. This directly counteracts repetitive output by encouraging the model to explore more diverse word choices, which is a standard technique for improving creativity in text generation.
Exam trap
Google Cloud often tests the misconception that decreasing sampling thresholds (like top-k or beam width) increases diversity, when in fact they reduce the candidate pool and make output more deterministic.
How to eliminate wrong answers
Option B is wrong because decreasing beam search width to 1 reduces the number of candidate sequences considered, which actually makes output more deterministic and less diverse, worsening repetitiveness. Option C is wrong because decreasing the top-k sampling threshold restricts the model to only the k most likely tokens, which reduces diversity and can increase repetition. Option D is wrong because adding more examples of repetitive text to the training dataset would reinforce the unwanted behavior, making the model more likely to generate repetitive output, not less.
A financial services company uses a generative AI model to produce customer-facing investment advice. They need to ensure the model's outputs can be traced back to specific sources. Which explainability technique is BEST suited for this requirement?
Grounding ensures every output can be traced to a specific source document or data point, meeting traceability needs.
Why this answer
Grounding (citing sources) is the best technique because it directly links each generated output to specific, verifiable source documents or data points. This ensures traceability, which is critical for regulated financial advice where every claim must be attributable to a known reference, such as a regulatory filing or market data feed.
Exam trap
The Generative AI Leader exam often tests the misconception that chain-of-thought reasoning provides traceability, but it only explains the model's reasoning path, not the external source of the information.
How to eliminate wrong answers
Option A is wrong because confidence indicators only provide a probabilistic score (e.g., 0.95) for the model's certainty, not a traceable source citation. Option B is wrong because Model Cards are documentation artifacts that describe a model's intended use, performance, and limitations, but they do not provide per-output source attribution. Option C is wrong because chain-of-thought reasoning reveals the model's internal step-by-step logic but does not cite external sources; it can even generate plausible but false reasoning paths (hallucinations) without grounding.
A team uses Vertex AI to host a large language model. They want to reduce latency for real-time applications. What is the best strategy?
Quantization reduces model size and speeds up inference.
Why this answer
Model quantization reduces the precision of the model's weights (e.g., from FP32 to INT8), which decreases memory footprint and computational requirements, directly lowering inference latency for real-time applications on Vertex AI. This is a standard optimization technique for deploying large language models with minimal accuracy loss while meeting latency SLAs.
Exam trap
Google Cloud often tests the misconception that scaling resources (replicas) directly reduces latency, when in fact latency optimization requires algorithmic changes like quantization or pruning, not just horizontal scaling.
How to eliminate wrong answers
Option A is wrong because increasing the number of replicas improves throughput and availability but does not reduce per-request latency; it may even add overhead from load balancing. Option B is wrong because switching to a smaller model reduces latency but sacrifices model capability and output quality, which is not a 'best' strategy when the team specifically needs a large language model. Option D is wrong because batch prediction is designed for asynchronous, high-throughput scenarios and introduces higher latency per request, making it unsuitable for real-time applications.
A company is comparing Google Cloud's AI offerings with AWS Bedrock and Azure OpenAI. They need to select a platform that provides: 1) native multimodal capabilities (image, video, text in one model), 2) integration with enterprise productivity tools, and 3) custom TPU hardware for training. Which THREE differentiators are unique to Google Cloud?
Google Cloud AI integrates seamlessly with Workspace.
Why this answer
Gemini's native multimodality, Workspace integration, and TPU infrastructure are unique to Google Cloud. AWS and Azure do not offer the same combination.
A team notices the RAG pipeline sometimes retrieves irrelevant documents. Which THREE improvements should they consider? (Choose three.)
Reranks retrieved documents by relevance.
Why this answer
Using a higher quality embedding model improves semantic understanding, adding a reranking step refines results, and reducing the number of retrieved documents reduces noise. Increasing chunk size can dilute relevance, and using exact keyword matching loses semantic context.
A large enterprise is migrating their on-premise ML workloads to Vertex AI. They have a custom PyTorch model for text classification that they want to serve with minimal code changes. Which Vertex AI capability should they use for model serving?
Custom containers support any framework and allow minimal code changes.
Why this answer
Vertex AI Prediction with a custom container allows the enterprise to package their existing PyTorch model with any custom dependencies or runtime configurations into a Docker container, enabling deployment with minimal code changes. This approach provides full control over the serving environment while leveraging Vertex AI's managed infrastructure for scaling, monitoring, and endpoint management.
Exam trap
The trap here is that candidates assume Vertex AI provides a pre-built PyTorch runtime similar to TensorFlow, but in reality, PyTorch models require a custom container because Vertex AI's managed runtimes only support TensorFlow, scikit-learn, and XGBoost natively.
How to eliminate wrong answers
Option A is wrong because Vertex AI Endpoints with a pre-built PyTorch runtime does not exist as a managed service; Vertex AI offers pre-built containers for TensorFlow, scikit-learn, and XGBoost, but not for PyTorch, so a custom container is required. Option C is wrong because Vertex AI Model Garden provides access to pre-trained foundation models and model hubs, not a mechanism for serving custom PyTorch models with minimal code changes. Option D is wrong because Vertex AI Vector Search is designed for approximate nearest neighbor (ANN) search on embeddings, not for serving text classification models directly.
A developer wants to quickly prototype a multimodal application that can process images and text using the Gemini API without incurring any cost. Which access tier should they use?
Google AI Studio provides a free tier for experimentation with Gemini API.
Why this answer
Google AI Studio (option D) is the correct choice because it provides a free, browser-based environment specifically designed for prototyping with the Gemini API, including multimodal capabilities for images and text, without requiring any payment or cloud billing setup. This makes it ideal for quick experimentation and cost-free development.
Exam trap
The trap here is that candidates often confuse Google AI Studio with Vertex AI Gemini API, assuming both require billing, or they overlook that Colab Enterprise and Cloud Shell are not designed for direct, cost-free Gemini API prototyping.
How to eliminate wrong answers
Option A is wrong because Colab Enterprise is a paid Google Cloud service that requires billing activation and is intended for production-grade notebook environments, not for free prototyping. Option B is wrong because Cloud Shell provides a command-line interface and basic cloud resources but does not natively offer the Gemini API or a multimodal prototyping interface without additional setup and potential costs. Option C is wrong because the Vertex AI Gemini API is a paid enterprise service that requires a Google Cloud project with billing enabled, and while it offers the same API, it is not cost-free for prototyping.
To ensure that a generative AI model uses the most current information from the web for answering user queries, which Vertex AI feature should be enabled?
Correct: This feature retrieves current web information to ground responses.
Why this answer
Grounding with Google Search is the correct feature because it enables the model to retrieve and reference real-time information from the web, ensuring responses are based on the most current data available. This is achieved by integrating Google Search results directly into the model's generation process, allowing it to cite live sources and reduce hallucinations from outdated training data.
Exam trap
Google Cloud often tests the distinction between features that improve output quality through external data retrieval (Grounding) versus those that modify the model's internal behavior (tuning, caching, filtering), leading candidates to confuse safety or optimization features with live data access.
How to eliminate wrong answers
Option B is wrong because safety filters are designed to block harmful or inappropriate content, not to fetch current web information. Option C is wrong because context caching stores frequently accessed context to reduce latency and cost, but it does not provide live web data. Option D is wrong because model tuning adjusts the model's parameters on a specific dataset to improve performance on a task, but it does not enable real-time web retrieval.
A company wants to measure the ROI of a GenAI-based report generation tool. Which metric is MOST directly tied to business value?
Time saved translates to cost savings and productivity, directly impacting ROI.
Why this answer
Time saved per report directly quantifies productivity gains, a key component of ROI. Other metrics are indirect or operational.
A developer is using Vertex AI Studio to test prompts for a text generation model. They want the model to follow a specific output format (JSON). Which prompt engineering approach is most effective?
Providing an example gives the model a concrete template to follow.
Why this answer
Including a few-shot example of the exact JSON format in the prompt provides the model with a concrete pattern to follow, which is the most reliable method for enforcing structured output in generative models. Few-shot prompting leverages in-context learning, where the model uses the provided example to infer the desired schema and formatting rules, reducing ambiguity and improving adherence to the specified JSON structure.
Exam trap
Google Cloud often tests the misconception that system instructions or hyperparameter tuning alone can enforce output format, when in practice, few-shot examples are the most direct and reliable method for guiding model behavior in structured generation tasks.
How to eliminate wrong answers
Option A is wrong because setting stop sequences to '}' would prematurely terminate generation at the first closing brace, which may cut off nested JSON objects or arrays, and does not guarantee the model outputs valid JSON from the start. Option C is wrong because a system instruction like 'Always output JSON' is a high-level directive that models often fail to follow precisely without explicit formatting examples, as they may still produce markdown, extra text, or malformed JSON. Option D is wrong because setting temperature to 0 makes output deterministic but does not enforce a specific output format; the model could still generate non-JSON text or deviate from the required schema, as temperature controls randomness, not structure.
A developer is building a code‑generation assistant using the Codey API on Vertex AI. The assistant should generate Python functions based on natural language descriptions. However, the generated code sometimes contains syntax errors. Which parameter adjustment would MOST directly help reduce syntax errors?
Lower temperature makes outputs more conservative and less random, reducing the likelihood of generating invalid syntax.
Why this answer
Reducing temperature makes the model more deterministic, which typically reduces creative but incorrect outputs like syntax errors. Prompt engineering can also help, but adjusting temperature is the simplest direct fix. Increasing max tokens or changing top-k does not directly address syntax correctness.
A company is evaluating Google Cloud vs AWS Bedrock for building a multimodal application that needs to understand images, video, and text in a single model. Which unique Google Cloud capability supports this requirement?
Gemini is designed for multimodal inputs in a single model.
Why this answer
Gemini is a natively multimodal model capable of processing text, images, audio, video, and code. AWS Bedrock's models are primarily text or image-only, lacking native video understanding.
A team is using Vertex AI Pipelines to deploy a generative AI model for real-time inference. The model sometimes generates harmful content. They want to implement a safety filter that checks the output before returning it to the user, but they need to minimize latency. Which approach best balances safety and performance?
A small classifier adds minimal latency while providing effective filtering.
Why this answer
Deploying a secondary lightweight classifier (e.g., a distilled BERT or a small logistic regression model) as a post-processing filter allows real-time inference with minimal latency overhead. This approach decouples safety from the primary generative model, enabling fast rejection of harmful outputs without retraining or blocking the main inference pipeline.
Exam trap
Google Cloud often tests the misconception that safety must be integrated into the generative model itself (e.g., via retraining or fine-tuning), when in practice a separate, lightweight post-processing filter is the standard for low-latency production systems.
How to eliminate wrong answers
Option B is wrong because retraining the model on every flagged harmful output is computationally expensive, introduces significant latency, and can lead to catastrophic forgetting or overfitting to specific examples, making it impractical for real-time inference. Option C is wrong because manual review of all outputs introduces unacceptable latency and does not scale, violating the requirement to minimize latency. Option D is wrong because disabling safety checks entirely eliminates the safety requirement, which is explicitly needed, and would expose users to harmful content, failing the core objective.
A financial services firm uses a fine-tuned Gemini model in Vertex AI for regulatory compliance checks. They notice that token usage is high, increasing costs. They want to reduce costs without sacrificing accuracy. Which approach should they take?
Directly reduces output tokens; precise prompts maintain accuracy.
Why this answer
Reducing max output tokens directly lowers the number of tokens generated per request, which is the primary cost driver in pay-per-token models like Gemini. Using more precise prompts further reduces token waste by guiding the model to produce concise, relevant outputs without sacrificing accuracy, as compliance checks often require specific, structured responses rather than verbose explanations.
Exam trap
The trap here is that candidates often confuse cost-reduction strategies that affect model behavior (like temperature or model size) with those that directly reduce token count, leading them to pick options that change output quality rather than token usage.
How to eliminate wrong answers
Option A is wrong because switching to a smaller base model like PaLM 2 Bison may reduce per-token cost but can degrade accuracy on complex regulatory compliance tasks, as smaller models have less capacity for nuanced understanding and may miss critical compliance nuances. Option B is wrong because context caching is designed to reduce latency and cost for repeated identical prompts by reusing cached responses, but it does not help when each compliance check involves unique input data (e.g., different contracts or transactions), making cache hits unlikely. Option D is wrong because setting temperature to 0.0 makes the model deterministic but does not reduce token usage; it may even increase token count if the model becomes overly repetitive or verbose in its attempts to be precise.
A research team uses a generative AI model to analyze historical texts. They want to provide users with insight into the model's reasoning process. Which explainability technique should they implement?
Chain-of-thought reveals the intermediate steps of reasoning, making the model's process transparent.
Why this answer
Chain-of-thought reasoning provides step-by-step explanations of how the model arrived at its conclusion, enhancing transparency.
An e-commerce company is using Vertex AI PaLM 2 for Text (via Model Garden) to generate product descriptions. They have an existing pipeline that calls the model with a prompt including product attributes. Recently, they migrated to the Gemini API. The team notices that the Gemini model sometimes outputs descriptions that are factually inconsistent with the input (e.g., wrong color or size). This was less frequent with PaLM 2. They have not changed the prompts. What is the most likely cause and solution?
Different models may require slight prompt adjustments; lower temperature and clearer instructions improve factual precision.
Why this answer
The core issue is that the prompt, originally optimized for PaLM 2, may not be sufficiently explicit for the Gemini model's different instruction-following behavior. By making the prompt more explicit about adhering strictly to the input data and reducing the temperature (e.g., to 0.2 or lower), the model's output becomes more deterministic and less prone to hallucinating incorrect attributes. This directly addresses the factual inconsistency without changing the model family, leveraging Gemini's ability to follow detailed instructions when properly guided.
Exam trap
The trap here is that candidates assume model migration is the root cause and choose to revert (Option A), when in fact the real issue is prompt adaptation and hyperparameter tuning for the new model's behavior.
How to eliminate wrong answers
Option A is wrong because reverting to PaLM 2 ignores the fact that the prompt was not optimized for Gemini; the issue is prompt engineering and hyperparameter tuning, not model reliability. Option B is wrong because negative prompts are not a standard mechanism in Gemini or PaLM 2 for text generation; they are used in image generation models (e.g., Imagen) to avoid certain concepts, not to enforce factual consistency in text outputs. Option D is wrong because increasing temperature would make outputs more random and less deterministic, worsening the factual inconsistency problem, not solving it.
A developer wants to generate Python code using Google Cloud's generative AI. Which model should they invoke?
Codey is designed for code generation.
Why this answer
Codey is Google Cloud's family of models specifically designed for code generation, completion, and chat, built on the PaLM 2 architecture and fine-tuned on code-heavy datasets. For a developer needing to generate Python code, Codey is the correct choice because it is purpose-built for code-related tasks, unlike other models that specialize in different modalities.
Exam trap
The trap here is that candidates may confuse Chirp (audio) or Imagen (image) with code generation because all are Google Cloud generative AI offerings, but each is specialized for a distinct modality, and the question explicitly asks for code generation.
How to eliminate wrong answers
Option A is wrong because Chirp is Google Cloud's speech-to-text model, designed for audio transcription, not code generation. Option C is wrong because Imagen is a text-to-image generation model, focused on creating visual content from text prompts, not code. Option D is wrong because Meena is a general-purpose conversational AI model (predecessor to LaMDA) optimized for open-domain dialogue, not for generating syntactically correct Python code.
An organization is planning to roll out a generative AI internal knowledge base assistant to employees. They want to ensure adoption and manage change effectively. Which two change management practices should they prioritize? (Choose TWO)
Training empowers users to get better results, increasing satisfaction and adoption.
Why this answer
Effective prompt engineering is critical for generative AI assistants; without training, employees may produce vague or poorly structured prompts, leading to irrelevant or low-quality responses, which undermines adoption. Providing training on prompt writing directly addresses the skill gap and empowers users to leverage the assistant effectively.
Exam trap
The Generative AI Leader exam often tests the distinction between 'maximizing immediate impact' (Option B) and 'phased adoption with feedback loops' (Option D), where candidates mistakenly choose a rapid full rollout thinking it drives faster adoption, ignoring the proven change management principle of starting small to build advocacy and refine the tool.
A media company wants to use a generative AI model to create marketing copy that includes citations to original sources. Which feature should they enable to ensure the model provides accurate attributions?
Grounding enables the model to reference and cite specific sources, ensuring attribution and factual accuracy.
Why this answer
Grounding allows the model to cite sources, improving explainability and trustworthiness by connecting outputs to verifiable information.
A social media company uses a generative AI to moderate user comments. They need to filter hate speech, violence, and sexual content. What is the most efficient way to implement content safety in Vertex AI?
Safety filters are designed to detect hate speech, violence, and sexual content out of the box.
Why this answer
Google's pre-built safety filters in Vertex AI are specifically designed for content moderation tasks like hate speech, violence, and sexual content detection. They are immediately available, require no custom training, and integrate directly with Vertex AI's generative AI workflows, making them the most efficient choice for a social media company needing rapid deployment.
Exam trap
The Generative AI Leader exam often tests the misconception that custom training (AutoML) is always better for domain-specific tasks, but here the pre-built filters are already optimized for the exact content categories needed, making custom training unnecessary and inefficient.
How to eliminate wrong answers
Option A is wrong because hiring human moderators is not efficient at scale; it introduces latency, high cost, and inconsistency, and does not leverage AI automation. Option B is wrong because using a third-party API introduces additional latency, cost, and potential data privacy concerns, and it does not integrate natively with Vertex AI's generative AI pipeline. Option C is wrong because training a custom content classifier from scratch using Vertex AI AutoML requires significant labeled data, time, and compute resources, which is inefficient compared to using pre-built, optimized safety filters.
A company is using a generative AI model for internal report generation. They notice costs are high because each request processes large amounts of text. Which business strategy would most effectively reduce costs while maintaining quality?
A smaller fine-tuned model can provide sufficient quality at lower cost for specific tasks.
Why this answer
Fine-tuning a smaller model on a specialized dataset reduces computational cost per inference because smaller models have fewer parameters and require less memory and processing power. By tailoring the model to the company's specific domain (e.g., internal reports), it can maintain output quality comparable to a larger general-purpose model, directly addressing the cost-per-request issue without sacrificing accuracy.
Exam trap
Google Cloud often tests the misconception that 'bigger is always better' or that caching universally reduces costs, but the trap here is that candidates overlook the unique nature of generative AI outputs and the cost benefits of model specialization over raw scale or caching.
How to eliminate wrong answers
Option B is wrong because using a more powerful model typically increases per-request cost and latency, and while it may reduce retries, the net cost often rises due to higher compute requirements. Option C is wrong because caching only helps if identical requests are repeated frequently; for generative AI report generation, each request is often unique (different text inputs), making caching ineffective for reducing per-request processing costs. Option D is wrong because increasing batch size for online predictions can reduce per-request cost only if requests are batched together, but online (real-time) predictions usually require low latency and process one request at a time, so larger batch sizes are not applicable and may increase latency.
A company with limited AI expertise wants to adopt gen AI. They need a solution that integrates with existing data and applications. Which Google Cloud offering is best?
Provides a low-code platform for building and deploying gen AI agents that integrate with enterprise data and applications.
Why this answer
Vertex AI Agent Builder is designed for building conversational AI agents with easy integration to enterprise data sources. Option A is wrong because Colab Enterprise is a notebook environment, not a full solution. Option B is wrong because Apigee is an API management platform.
Option C is wrong because BigQuery ML is for SQL-based ML, not gen AI agents.
A company is deploying a generative AI application using Vertex AI. They need to minimize latency for real‑time inference while maintaining high quality. Which TWO actions are most effective?
Flash is optimized for lower latency.
Why this answer
Using a smaller model (e.g., Gemini Flash) reduces latency, and reducing the max output tokens limits generation time. Increasing temperature does not affect latency; using a larger model increases latency; batching is for throughput, not single‑request latency.
A company wants to adopt GenAI for internal knowledge base Q&A. They have a collection of PDFs, internal wikis, and slide decks. Which TWO services should they combine to build a RAG-based solution?
Grounding with Vertex AI Search retrieves relevant document chunks to augment LLM responses.
Why this answer
Vertex AI Agent Builder can create a data store from these documents, and grounding with Vertex AI Search retrieves relevant chunks for the LLM.
A data science team wants to decide between using a pre-built API (e.g., Vertex AI Gemini API) and fine-tuning a custom model for a specific business task. Which TWO factors are most important in making this build versus buy decision?
Cost comparison is crucial; fine-tuning may reduce per-query cost for high volume.
Why this answer
Pre-built APIs are best when task complexity is low and data privacy is not a concern. Fine-tuning is better for high complexity and when data must remain private.
A global nonprofit organization is deploying a generative AI chatbot to provide educational content in multiple languages to underserved communities. They operate in regions with limited internet connectivity. The chatbot must work offline or with minimal data usage. The team has a moderate budget and limited technical staff. Which deployment strategy should they use?
Enables offline inference with low resource usage.
Why this answer
Deploying a distilled version of the model on edge devices using TensorFlow Lite directly addresses the constraints of offline operation, minimal data usage, and limited technical staff. Distillation reduces model size and computational requirements, enabling inference on local hardware without cloud dependency, which is critical for underserved regions with intermittent connectivity.
Exam trap
The trap here is that candidates confuse 'distillation on edge' with 'distillation on cloud VMs' (Option D), overlooking that edge deployment is the only way to guarantee offline functionality, while cloud VMs still require network access for inference.
How to eliminate wrong answers
Option A is wrong because hosting a fine-tuned model on a cloud VM with auto-scaling requires constant internet connectivity for the chatbot to function, which fails the offline requirement. Option C is wrong because using a large foundation model via API requests from a mobile app incurs high data usage and relies on continuous cloud access, contradicting the need for minimal data usage and offline capability. Option D is wrong because deploying a distilled model on Google Cloud VM instances still requires internet connectivity for inference, missing the offline requirement, and does not leverage edge deployment for local processing.
A data scientist wants to apply reinforcement learning from human feedback (RLHF) to improve a chatbot's helpfulness. Which TWO steps are part of the RLHF process? (Select 2)
Human feedback is used to train a reward model.
Why this answer
RLHF typically involves collecting human rankings of model outputs and then training a reward model to score outputs, which is used to fine-tune the model via PPO.
A company is piloting a GenAI feature for internal knowledge base search. During the pilot, users report that the AI sometimes gives incorrect answers based on outdated documents. What is the MOST effective way to address this issue?
RAG retrieves the most relevant and current documents at inference time, directly addressing the issue of outdated information.
Why this answer
Grounding with Google Search (or enterprise search) ties responses to fresh sources, but ensuring the knowledge base is up-to-date and using RAG with the latest documents is more direct. Fine-tuning on old data will not fix freshness. Reducing temperature helps but does not solve outdated content.
Prompting to refuse uncertain answers can help but not as reliable as RAG.
A healthcare startup is using a large language model (LLM) to generate discharge summaries. To comply with regulations, they need to ensure that a human reviews all AI-generated summaries before they are sent to patients. Which Google Cloud feature should they use to enforce this workflow?
HITL enables configurable human review steps, ensuring every generated summary is checked by a qualified reviewer before release.
Why this answer
Human oversight is a key requirement for high-stakes AI systems. Vertex AI Human-in-the-Loop (HITL) provides a managed workflow to route predictions for human review, approval, or override before final output.
A company using Vertex AI for model training wants to reduce costs by automatically stopping training when the model's performance stops improving. Which Vertex AI feature should they configure?
Early stops training when a metric stops improving, saving time and cost.
Why this answer
Early stopping in Vertex AI Training stops training runs when performance plateaus. Hyperparameter tuning optimizes parameters, not cost. Workbench is for notebooks.
AutoML handles this automatically but is for specific model types.
A developer wants to use a pre-trained model to identify objects in images. Which Google Cloud AI API should they use?
Vision AI offers pre-trained models for object detection in images.
Why this answer
Vision AI provides pre-trained models for object detection, image classification, etc. Natural Language AI is for text, Speech-to-Text for audio, and Translation for text translation.
Which of the following is a key consideration when selecting a GenAI model for a cost-sensitive application?
Latency and throughput directly determine the infrastructure needed and thus the cost per inference.
Why this answer
For cost-sensitive applications, latency and throughput requirements directly impact infrastructure costs, as lower latency often requires more expensive compute resources (e.g., higher GPU memory, faster inference hardware) and higher throughput may necessitate scaling out instances. Model size in parameters is a secondary factor that influences latency and throughput, but the primary cost driver is the operational performance needed to meet service-level agreements (SLAs).
Exam trap
Google Cloud often tests the misconception that model size (parameters) is the primary cost driver, but the exam emphasizes that operational metrics like latency and throughput are the direct determinants of infrastructure cost in production.
How to eliminate wrong answers
Option A is wrong because model size in parameters affects memory and compute requirements but is not the key consideration for cost sensitivity; a smaller model can still be costly if latency or throughput demands are high. Option C is wrong because number of training epochs is a training-time hyperparameter that does not directly influence inference cost or operational cost in a deployed application. Option D is wrong because the model's training data source impacts bias, accuracy, and compliance, but not the direct operational cost of running inference at scale.
A retail company uses Vertex AI Agent Builder to create a virtual assistant for order tracking. Users frequently ask about delivery dates, but the assistant sometimes gives incorrect information. The team wants to improve accuracy without retraining the underlying model. Which technique should they apply?
Grounding allows the agent to retrieve up-to-date information from search or enterprise databases, reducing hallucinations about order status.
Why this answer
Grounding with Google Search or enterprise data sources (like order databases) ensures the agent retrieves real-time, accurate information instead of relying solely on the model's training data.
A financial services firm needs to use Gemini for analyzing customer transaction data. They require that all data remain within their VPC and that model inference logs be auditable. Which access tier should they choose?
Vertex AI offers VPC Service Controls, data isolation, and audit logging for enterprise compliance.
Why this answer
Vertex AI provides enterprise controls like VPC-SC, data isolation, and audit logging, while Google AI Studio is a prototyping environment without these guarantees.
A product manager wants to transparently communicate the capabilities and limitations of their new generative AI feature to stakeholders. Which THREE resources from the Responsible AI Toolkit should they use?
Model Cards provide structured information about model capabilities and limitations.
Why this answer
Model Cards document model details, Datasheets for Datasets document data, and the People + AI Guidebook provides best practices for communication. PAIR Explorables are for education, not documentation.
A developer is using the Gemini API to generate creative product taglines. The taglines are often bland and uncreative. The developer wants more variety and novelty in the outputs. Which parameter adjustment would most effectively increase the diversity of the generated taglines?
Higher temperature increases randomness, leading to more diverse and creative outputs.
Why this answer
Increasing temperature from 0.2 to 0.9 raises the randomness of token sampling, which directly increases the diversity and novelty of generated text. A low temperature (e.g., 0.2) makes the model highly deterministic, always picking the most probable next token, leading to bland outputs. A higher temperature (e.g., 0.9) allows less probable tokens to be selected more often, producing more creative and varied taglines.
Exam trap
The trap here is that candidates often confuse temperature with top_p, incorrectly assuming that lowering top_p increases diversity, when in fact it restricts the token pool and reduces variety.
How to eliminate wrong answers
Option A is wrong because decreasing top_p from 1.0 to 0.5 reduces the nucleus of tokens considered for sampling, which actually decreases diversity by cutting off the long tail of less probable tokens. Option B is wrong because setting frequency_penalty to 2.0 penalizes token repetition too aggressively, which can suppress natural language patterns and may reduce overall output quality without directly increasing novelty. Option D is wrong because decreasing temperature from 0.7 to 0.2 makes the model more deterministic, reducing randomness and thus decreasing diversity, which is the opposite of what the developer wants.
A research organization is building a generative AI model to assist in drug discovery by generating molecular structures. They have a large dataset of proprietary chemical compounds and want to train a model from scratch. They have extensive ML expertise but limited GPU resources. The organization must comply with strict data privacy regulations that prohibit data from leaving their on-premises environment. Which strategy enables them to train the model efficiently while meeting compliance?
Confidential VMs encrypt data in use, meeting privacy needs with scalable GPUs.
Why this answer
Google Cloud Confidential VMs with attached GPUs provide hardware-based memory encryption (using AMD SEV or Intel TDX) that protects data in use, enabling secure training on sensitive proprietary chemical data in the cloud. This allows the organization to leverage scalable GPU resources for efficient model training while maintaining compliance with strict data privacy regulations that prohibit data from leaving their on-premises environment.
Exam trap
Google Cloud often tests the misconception that any cloud GPU instance is sufficient for compliance, but the trap here is that standard GPU instances lack in-use memory encryption, which is required when data privacy regulations prohibit data from leaving the on-premises environment.
How to eliminate wrong answers
Option A is wrong because training entirely on-premises using existing servers would be inefficient due to limited GPU resources, leading to excessively long training times for a generative AI model from scratch. Option C is wrong because partnering with a cloud provider without specifying a secure, encrypted compute environment (like Confidential VMs) would expose the proprietary data to potential privacy risks and violate compliance requirements. Option D is wrong because transferring data to Google Cloud and using standard GPU instances does not provide the necessary in-use data encryption, leaving the data vulnerable during processing and failing to meet strict data privacy regulations.
A company deploys a Gemini model on Vertex AI for a healthcare application. They need to ensure that the model does not generate medical advice and that responses are grounded in trusted medical sources. Which combination of safety measures should they implement?
This combination ensures safety and factual grounding.
Why this answer
It combines two essential safety layers: safety filters block harmful content (including medical advice), and Vertex AI Grounding anchors responses to a labeled medical dataset, ensuring factual accuracy and compliance with healthcare regulations. This dual approach prevents the model from generating unverified or dangerous medical information while maintaining relevance to trusted sources.
Exam trap
The trap here is that candidates assume fine-tuning alone is sufficient for domain-specific safety, but without grounding and safety filters, the model can still hallucinate or generate unverified medical advice, which is a key distinction Google Cloud tests in the Generative AI Leader exam.
How to eliminate wrong answers
Option B is wrong because using a public dataset for grounding introduces unverified or non-authoritative medical information, and disabling safety filters removes the critical barrier against generating harmful or unlicensed medical advice. Option C is wrong because safety filters alone cannot ensure responses are grounded in trusted medical sources; they only block explicit content but do not prevent the model from fabricating medical facts. Option D is wrong because fine-tuning on a curated dataset does not guarantee real-time grounding in trusted sources, and disabling safety filters exposes the application to generating unverified medical advice, which is unacceptable in healthcare.
A company wants to build a chatbot using Vertex AI that can answer customer questions based on their internal knowledge base. Which Google Cloud service should they use to store and retrieve the knowledge base efficiently?
Vertex AI Vector Search provides scalable vector similarity search for knowledge retrieval.
Why this answer
Vertex AI Vector Search is the correct choice because it is purpose-built for semantic similarity search over embeddings, enabling the chatbot to retrieve relevant chunks from the knowledge base based on meaning rather than exact keyword matches. It integrates natively with Vertex AI and supports high-dimensional vector indexing, making it efficient for large-scale retrieval-augmented generation (RAG) workflows.
Exam trap
The trap here is that Google Cloud often tests the rebranding of Vertex AI Matching Engine to Vertex AI Vector Search, leading candidates to select the outdated service name (Matching Engine) instead of the current correct name (Vector Search).
How to eliminate wrong answers
Option A is wrong because Cloud Storage is an object storage service for unstructured data, not a vector database; it lacks built-in similarity search capabilities and would require additional services to perform semantic retrieval. Option C is wrong because BigQuery is a serverless data warehouse designed for SQL-based analytics on structured data, not for storing and querying dense vector embeddings with approximate nearest neighbor (ANN) search. Option D is wrong because Vertex AI Matching Engine is the previous name for what is now Vertex AI Vector Search; the service was rebranded, so the current correct name is Vector Search, making Matching Engine a deprecated or legacy term in this context.
What is the primary function of embeddings in the context of generative AI?
Embeddings map words, sentences, or images into vectors that represent semantic meaning, used for similarity search and retrieval.
Why this answer
Embeddings are numerical representations of data that capture semantic meaning, enabling vector search and semantic similarity. They are not used for token generation directly, and they are not the same as temperature or context window.
A project manager wants to reduce the cost of using Gemini API for batch processing of customer feedback. The team is on a tight budget. Which cost management strategy is MOST effective?
Batch requests are cheaper than individual calls for large volumes.
Why this answer
Batch requests reduce per-token cost by grouping multiple prompts into one API call. Caching helps but is less impactful for varied feedback.
A machine learning engineer is tuning a large language model on Vertex AI for question answering. They want to evaluate the model's performance before deployment. Which THREE metrics should they consider?
F1 balances precision and recall.
Why this answer
The F1 score is a key metric for evaluating question-answering models because it measures the harmonic mean of precision and recall, capturing how well the model's predicted answer overlaps with the ground truth answer. In the context of Vertex AI and large language models, F1 score is particularly useful when answers are not exact matches but contain partial correct information, making it a standard evaluation metric for extractive and generative QA tasks.
Exam trap
The trap here is that candidates confuse operational metrics (like cost or training time) with evaluation metrics that directly measure model output quality, leading them to select options that are irrelevant to performance assessment.
An organization wants to use AI-generated images commercially. According to Google's AI principles and copyright guidelines, what should they do FIRST?
Understanding the data used to train the model helps assess copyright risks.
Why this answer
The primary risk is copyright infringement from training data. Verifying that the model's training data was properly licensed is essential before using outputs commercially.
A company develops a generative AI model for resume screening. They discover that the model is rejecting candidates from certain demographic groups disproportionately. Which step should they take first to address unfair bias?
Auditing the data is the first step to identify if bias stems from under- or over-representation of groups.
Why this answer
Before attempting to fix bias, it is essential to evaluate the training data for representativeness. Bias often originates from imbalanced or unrepresentative training data.
A team is deciding between using fine-tuning and in-context learning for a document classification task. They have 500 labeled examples and need low latency. Which TWO statements are accurate?
With 500 examples, fine-tuning can adapt the model better to the task, often yielding higher accuracy.
Why this answer
Fine-tuning updates the model's weights on a labeled dataset, which generally leads to higher accuracy than in-context learning when sufficient labeled data (like 500 examples) is available. In-context learning relies on the model's pre-existing knowledge and a few examples in the prompt, which often yields lower accuracy for complex classification tasks.
Exam trap
Google often tests the misconception that in-context learning always has lower latency than fine-tuning, but the trap is that latency depends on prompt length and model architecture, not just the absence of a training step.
Which Google Cloud offering allows you to create a machine-readable document that describes a model's intended use, performance, and limitations?
Model Cards are the standard for documenting model details.
Why this answer
Model Cards are designed to provide transparent documentation of model behavior, including intended use, performance, and limitations.
A company wants to build a multimodal application that can analyze images and generate captions. They are considering using Google Cloud AI. Which TWO services can be directly used for this purpose?
Vision AI can analyze images to extract features and labels.
Why this answer
Vision AI (option A) is correct because it provides pre-trained models for image analysis, including object detection and image labeling, which are essential for generating captions. The Gemini API (option E) is correct because it is a multimodal model that can directly process images and text, enabling it to analyze images and produce descriptive captions without needing separate services.
Exam trap
The trap here is that candidates may mistakenly think Natural Language AI (option C) can generate captions because it handles text, but it cannot process images, while the Gemini API's multimodal capability is often overlooked in favor of more familiar single-purpose services.
An organization uses a fine-tuned model for medical diagnosis and must comply with HIPAA. Which measure is essential when deploying the model on Vertex AI?
This ensures the endpoint is not publicly accessible, a key requirement for HIPAA.
Why this answer
HIPAA requires that patient data be protected from unauthorized access during transmission and deployment. Using a private Google Cloud Access endpoint with external internet access disabled ensures that the model endpoint is only reachable within the organization's VPC network, preventing data exposure over the public internet and meeting HIPAA's security rule for safeguarding electronic protected health information (ePHI).
Exam trap
Google Cloud often tests the misconception that encryption at rest or basic data storage features are sufficient for HIPAA compliance, when in fact network-level access controls (like private endpoints) are the critical measure for protecting ePHI during model inference.
How to eliminate wrong answers
Option A is wrong because storing patient data in Cloud Storage with object versioning provides data retention and recovery capabilities but does not address the core requirement of securing the model endpoint or controlling network access, which is essential for HIPAA compliance. Option B is wrong because enabling encryption at rest for all resources is a baseline security practice and is already enabled by default in Google Cloud; it does not specifically address the need to restrict network access to the deployed model endpoint, which is a key HIPAA requirement. Option C is wrong because using a publicly accessible endpoint for faster response times directly violates HIPAA's requirement to protect ePHI from unauthorized access, as a public endpoint exposes the model to the internet and increases the risk of data breaches.
A developer wants to integrate Gemini multimodal capabilities (text + image) into a mobile app using Python. Which Google Cloud client library should they use?
The Vertex AI client library supports Gemini API for multimodal generation.
Why this answer
The Vertex AI client library (google-cloud-aiplatform) provides the Generative AI SDK that supports multimodal capabilities, including the ability to send both text and image inputs to Gemini models. This library directly exposes the `GenerativeModel` class with methods like `generate_content()` that accept `Part` objects containing image data (e.g., `Part.from_image()` or `Part.from_uri()`), making it the correct choice for integrating Gemini multimodal features into a Python mobile app backend.
Exam trap
The trap here is that candidates confuse specialized single-modality APIs (Vision, Natural Language) with the unified multimodal API provided by Vertex AI, assuming that combining separate services is equivalent to Gemini's native multimodal reasoning.
How to eliminate wrong answers
Option A is wrong because Dialogflow CX is a conversational AI platform for building chatbots and virtual agents, not a library for directly accessing Gemini multimodal models; it lacks the low-level API to construct multimodal requests with image parts. Option C is wrong because Cloud Vision API is a specialized service for image analysis (e.g., object detection, OCR) and does not provide access to Gemini's generative multimodal capabilities or its text+image reasoning. Option D is wrong because Natural Language API is designed for text-only analysis (e.g., sentiment, entity extraction) and cannot process image inputs or generate multimodal responses.
A company uses Gemini 1.5 Pro to analyze customer call transcripts and generate summaries. They notice that the summaries occasionally include fabricated details that were not in the transcript. Which technique is specifically designed to reduce such hallucinations?
RAG retrieves relevant transcript segments and provides them to the model as context, grounding the output in actual data and reducing hallucinations.
Why this answer
Grounding with RAG retrieves factual data from trusted sources to condition the model's response, directly reducing hallucinations. Prompt engineering can help but does not guarantee factual accuracy; fine-tuning may reduce but not eliminate; temperature reduction makes output more deterministic but does not address factuality from external data.
Refer to the exhibit. A team has deployed a model to an endpoint with the configuration shown. They notice that during peak traffic, the endpoint frequently returns 429 (Too Many Requests) errors. Which action should they take to resolve this issue?
More minimum replicas provide capacity for sudden traffic spikes.
Why this answer
Increasing MIN_REPLICA_COUNT ensures a minimum number of replicas are always available to handle traffic bursts, reducing 429 errors. Other options would not help or would worsen the problem.
You are using Vertex AI Model Garden to deploy a Llama model. Which deployment option provides the best latency for real-time inference?
Fixed replicas ensure always-on instances for low latency.
Why this answer
Deploying a Llama model to a Vertex AI Endpoint with a fixed number of replicas ensures that compute resources are pre-allocated and always warm, minimizing cold-start latency and providing consistent, low-latency responses for real-time inference. This approach uses a dedicated endpoint with persistent instances, which is optimized for sub-second response times required by interactive applications.
Exam trap
The trap here is that candidates confuse 'autoscaling' (which optimizes cost) with 'low latency' (which requires pre-provisioned resources), and they overlook that MaaS with autoscaling introduces cold-start delays that are unacceptable for real-time inference.
How to eliminate wrong answers
Option A is wrong because Batch Prediction is designed for asynchronous, high-throughput processing of large datasets and does not provide real-time inference; it queues jobs and returns results later, making it unsuitable for latency-sensitive applications. Option B is wrong because deploying to a Compute Engine VM requires manual management of the serving infrastructure, lacks built-in autoscaling, load balancing, and health checks that Vertex AI Endpoints provide, leading to higher operational overhead and potential latency spikes under variable load. Option D is wrong because MaaS (Model-as-a-Service) with autoscaling introduces cold-start delays when scaling from zero replicas, and autoscaling policies may not react quickly enough to sudden traffic bursts, causing increased latency during scale-up events.
A financial services company wants to automate contract analysis to extract key clauses and identify risky terms. They have thousands of PDF contracts and need a solution that can be quickly integrated into their existing document management system. Which Google Cloud service is MOST suitable?
DocAI provides pre-trained models for parsing contracts and extracting custom entities, directly integrated via API.
Why this answer
DocAI is purpose-built for document understanding and structured extraction from PDFs. Vertex AI Studio focuses on prompt design for generative models, while Model Garden is for selecting models. Vertex AI Agent Builder is for building conversational agents, not document processing.
Practice Generative AI Leader by domain
Target a specific domain to shore up weak areas.