Google Generative-AI-Leader Practice Test Questions Answers

The most direct and effective way to influence the style, personality, and knowledge context of an AI ' s response is through Role Prompting.

Role Prompting involves instructing the model to assume a specific persona (a " role " ) before responding. By assigning the AI the role of an " experienced customer service representative " (B), the model is implicitly directed to adopt a professional, helpful, and empathetic tone. Furthermore, specifying " with corporate knowledge " directs the model to prioritize responses consistent with internal company policies. This technique is a foundational element of prompt engineering, often used in conjunction with other methods (like grounding, if specific policy documents were needed) to dramatically shift the output style and relevance.

While Few-shot prompting (D) could provide examples to influence style, it ' s less efficient than a clear role instruction and still requires the model to infer the persona. Prompt Chaining (A) is used to manage multi-turn conversation memory, not to set the tone or persona. Therefore, defining the Role is the core technique for establishing both the desired tone and the necessary professional context in a single instruction.

(Reference: Google ' s documentation on prompt engineering for customer service shows examples where users begin the prompt with " I am a customer service representative " to set the tone and persona for the generated response, confirming Role Prompting as the technique for ensuring style and consistency.)

Secure-by-design infrastructure means that security protections are incorporated into Google Cloud’s hardware, network, and software architecture from their initial design rather than being added only after deployment. Google’s infrastructure applies layered controls such as custom hardware security, encryption, identity-based access, network protections, continuous monitoring, and a hardened global architecture. These measures protect workloads and data at multiple levels and provide a secure foundation for generative AI applications. Network protection describes only one part of this broader architecture. Operations management concerns administration and reliability processes, while security oversight refers generally to governance and supervision rather than inherent technical controls. Because the organization is asking for the term covering security measures fundamentally integrated into Google’s global network and hardware, secure-by-design infrastructure is the precise and comprehensive answer.

Supervised machine learning trains a model using examples containing both input features and known output labels. Historical support tickets can include characteristics such as issue category, priority, customer type, and complexity, together with the actual resolution time. The model learns the relationship between those labeled examples and then predicts resolution time for new tickets. The remaining scenarios focus on discovering previously unknown structures in unlabeled data. Identifying customer segments from purchase history and finding behavior clusters from website clicks are clustering tasks. Automatically discovering recurring topics in customer reviews is generally a topic-modeling task. These are examples of unsupervised learning because the system is not trained against predetermined target outcomes. Therefore, predicting resolution time from past support tickets with known resolution-time labels is the clear example of supervised machine learning.

First-party information is data owned or directly collected by an organization, such as customer transactions, account records, support history, and purchase details. A chatbot cannot reliably answer a question about a customer’s recent purchase history from a foundation model’s general training data. It must be grounded in the company’s current, authorized customer records to provide an accurate and personalized response. Appropriate identity verification and access controls must also be applied before retrieving the information. Public sentiment is generally evaluated using external public data, while definitions of common scientific terms can normally be answered from general model knowledge. Summarizing global news requires grounding in external news sources rather than proprietary first-party records. Therefore, retrieving specific purchase-history information is the clearest situation where first-party grounding provides essential factual context.

Extensions provide a standardized mechanism through which an AI agent can invoke external APIs and complete real-world actions. A travel extension can connect the agent to flight and hotel services, pass structured parameters, retrieve current availability and prices, and submit reservations. This is more maintainable and consistent than embedding separate custom functions directly inside the agent. Pre-loaded data stores cannot satisfy the requirement because availability and pricing change continuously, while reservations require transactional access to the provider’s live system. “Plugins” is not the recommended Google Cloud mechanism identified for this agent-integration scenario. Extensions enable agents to move beyond answering questions by securely communicating with external services through defined interfaces. Consequently, they are the simplest and most effective Google-recommended method for integrating the travel agent with provider APIs.