Generative AI Leader Exam Questions and Answers

Enable Vertex AI Grounding with a custom enterprise data store containing the documentation.

Grounding with a custom enterprise data store retrieves passages from the product documentation and injects them into the prompt, so responses cite actual content rather than relying on parametric memory. This directly addresses the hallucination constraint in the stem.

Reduce the model size or add dropout regularization.

Adding dropout regularisation directly counteracts the overfitting causing validation loss to diverge from training loss. Reducing model capacity likewise limits memorisation of the custom fine-tuning data. Both address the generalisation gap the stem describes, where training loss falls while validation loss rises, restoring alignment between the two curves.

A text generation model (e.g., text-bison@001).

A text generation model such as text-bison@001 maps a feature list to fluent prose, satisfying the stem's requirement to produce product descriptions. Unlike classification or embedding models, it outputs free-form natural language, which is precisely the generative capability Vertex AI needs here.

They are typically pre-trained on large datasets.

Pre-training on large datasets is the defining characteristic of generative models: self-supervised learning over billions of tokens or images builds the broad statistical patterns later fine-tuned for specific tasks. This satisfies the stem's requirement for a true statement about how such models are typically built.

They can generate new content not seen in training.

Generative models produce novel outputs by sampling from learned probability distributions rather than retrieving stored examples, so responses are synthesised combinations not present verbatim in training data. This directly satisfies the stem's requirement for a true statement about generative capability.

Ensure the training data is diverse and representative.

Diverse, representative training data reduces demographic bias, preventing skewed diagnostic suggestions for under-represented patient groups. This satisfies the responsible AI requirement by addressing fairness at the data layer, where bias originates before model training begins.

Implement human oversight for all diagnostic suggestions.

Human oversight ensures a qualified clinician reviews every AI diagnostic suggestion before it affects patient care, catching errors and preserving clinical accountability. This satisfies the responsible AI requirement for meaningful human control in high-stakes medical decisions.

Provide clear disclaimers about the model's limitations.

Clear disclaimers communicate that model outputs are decision support, not confirmed diagnoses, so clinicians and patients interpret suggestions with appropriate caution. This satisfies the responsible AI requirement for transparency about capability limits in medical contexts.

Reduce the batch size

Out-of-memory during fine-tuning stems from the activation and gradient tensors held per training step. Reducing the batch size shrinks those tensors proportionally, lowering peak GPU memory below the accelerator's limit so the Vertex AI job can complete without changing the model architecture.

Implement a human-in-the-loop review process with clear accountability for AI-generated recommendations.

Human-in-the-loop review keeps a licensed clinician as the accountable decision-maker, so AI output informs rather than determines care. This satisfies the stem's regulatory and liability constraints: HIPAA compliance and malpractice exposure remain with the clinician, and Microsoft Entra ID can enforce role-based access to the review workflow.

The solution requires deep integration with proprietary data and unique domain-specific outputs.

Building from scratch allows fine-tuning or training on proprietary datasets, giving the model access to unique domain vocabulary and logic that a generic pre-built API cannot replicate. This directly satisfies the stem's constraint of deep proprietary-data integration and unique domain-specific outputs.

Implement a human review layer for high-risk topics and use automated fact-checking for all content, with a feedback loop for model improvement.

A human review layer catches high-risk factual and bias errors that automation misses, while automated fact-checking scales across all content. The feedback loop retrains the model, reducing recurrence without suppressing the personalisation that sustains user engagement.

Use a retrieval-augmented generation (RAG) system that queries a live database of sensor configurations.

RAG retrieves current sensor configurations from the live database at inference time, grounding each generated manual in up-to-date equipment state. Unlike fine-tuning, which bakes in stale weights, retrieval guarantees freshness and consistency with the latest configurations.

Establish a human-in-the-loop escalation process for sensitive interactions.

Sensitive or ambiguous queries can produce harmful or off-brand replies, so routing them to a human before the response reaches the customer prevents reputational damage. This satisfies the brand-reputation risk constraint by keeping a person accountable for high-stakes interactions.

Implement automated monitoring for toxic or off-brand language.

Automated monitoring continuously scans generated responses for toxic, biased or off-brand language, catching reputational risks at scale before they reach customers. This satisfies the brand-reputation constraint by enabling rapid detection and remediation rather than relying on manual review alone.

Develop region-specific prompt templates that incorporate local cultural references and legal requirements.

Region-specific prompt templates embed local cultural references and legal requirements into generation, ensuring each market's descriptions respect nuances while a shared template structure maintains cross-market consistency. This directly satisfies both the consistency and cultural-respect constraints in the stem.

Engage local marketing teams to review and approve AI-generated descriptions before publication.

Local marketing teams provide the cultural nuance that a single global prompt cannot encode, catching idioms, taboos and tone that literal translation misses. Their review-and-approval step enforces consistency across markets while adapting each description locally, directly satisfying the stem's dual requirement of uniformity and cultural sensitivity.

Use A/B testing to measure engagement metrics per region and iterate on prompts.

A/B testing per region supplies the empirical feedback loop the stem demands: engagement metrics reveal whether a prompt's tone or idiom lands culturally, letting the team iterate prompts rather than assume one wording suits every market. This directly satisfies the consistency-with-cultural-nuance constraint.

Vertex AI Search with document grounding

Vertex AI Search with document grounding indexes the Cloud Storage documents and serves grounded answers from an EU multi-region endpoint, keeping data resident in the EU while avoiding cross-region retrieval latency. This satisfies both the data residency and latency constraints.

Vertex AI Studio

Vertex AI Studio provides a no-code console for prompting and testing Gemini models directly, letting non-specialists generate product descriptions from keywords without building pipelines. This satisfies the startup's lack of ML experience and fastest time-to-market constraint, since no model training or infrastructure setup is required.

Vertex AI Agent Builder

Vertex AI Agent Builder orchestrates conversational agents that call backend systems through tools and extensions, satisfying the integration constraint for returns and order status. Unlike plain generative endpoints, it manages dialogue state and grounding, letting the chatbot retrieve live order data and execute return workflows rather than only answering static FAQs.

Use Vertex AI grounding with a knowledge base of verified entities

Grounding with a verified entity knowledge base constrains generation to retrieved, authoritative facts, so named entities come from the knowledge base rather than the model's parametric memory. This directly targets the factual-error constraint in the stem, reducing hallucinated actor names without retraining or prompt engineering.

Adjust safety settings to block medium-probability harmful content

Safety settings operate on probability thresholds per harm category, so lowering the block threshold from high to medium catches harmful content the model would otherwise emit. This tightens filtering beyond the existing high-probability configuration, directly reducing offensive outputs from the Gemini API.

Use Vertex AI Translation with custom glossaries

Custom glossaries let Vertex AI Translation enforce consistent rendering of domain-specific terms such as 'size chart' and 'return policy' across all 50 languages. Generic translation models may mistranslate these, so the glossary constraint is what preserves accuracy at scale.

Apply reinforcement learning from human feedback (RLHF)

Reinforcement learning from human feedback directly optimises the model against human preference rankings, aligning outputs with helpfulness rather than politeness alone. This satisfies the stem's constraint of improving response quality without sacrificing helpfulness, since reward signals penalise unhelpful replies and reinforce substantive ones during fine-tuning.

Implement constrained decoding with grammar rules

Constrained decoding masks tokens that would violate the target language's grammar at each generation step, so syntactically invalid sequences are never emitted. This directly reduces syntax errors at inference time, satisfying the stem's constraint of no retraining. Unlike fine-tuning, it requires no weight updates, only a grammar specification during decoding.

Fine-tune the model on a curated dataset of human faces

Fine-tuning adjusts the model's weights on curated face data, teaching it the specific facial structures and proportions it currently renders poorly. This directly targets the distorted-face failure mode, unlike prompt engineering or higher resolution, which cannot correct learned representation gaps.

Design a prompt that explicitly lists required sections

Explicitly listing the required sections in the prompt constrains the model's output structure, directing attention to mandatory clauses that free-form summarisation tends to drop. This prompt-engineering approach improves recall of critical content without retraining or architectural changes.

Fine-tune with a dataset of non-toxic responses and use RLHF

Supervised fine-tuning on non-toxic responses teaches safer output distributions, and RLHF then optimises the policy against a reward model balancing toxicity reduction with helpfulness. This directly targets the constraint of lowering toxicity without materially degrading response quality.

Apply a repetition penalty of 1.2

A repetition penalty of 1.2 directly down-weights tokens already generated, lowering their probability at each subsequent step. This penalises the excessive phrase recurrence the team observed, satisfying the requirement to reduce repetition without altering the model's weights or retraining.