CCNA Fundamentals of Generative AI Questions

The increasing validation loss while training loss decreases is a classic sign of overfitting, where the model memorizes the training data but fails to generalize. Reducing model size or adding dropout regularization directly combats overfitting by limiting the model's capacity or introducing noise during training, which forces the model to learn more robust features. This is the best course of action because it addresses the root cause without further exacerbating the problem.

Vertex AI Generative AI Studio is the correct answer because it is the Google Cloud service specifically designed to provide access to pre-trained foundation models like Gemini, allowing users to test, customize, and deploy them via a managed interface. Unlike other services, Generative AI Studio directly integrates with Gemini's API and offers prompt engineering, tuning, and model evaluation capabilities.

Temperature is the parameter that directly controls the randomness of the model's output. Lowering the temperature makes the model more likely to choose high-probability tokens, resulting in more deterministic and consistent text. This is exactly what the marketing team needs for focused product descriptions.

Expected Calibration Error (ECE) directly measures the alignment between a model's predicted confidence and its actual accuracy. In this scenario, high confidence on incorrect answers indicates miscalibration, and ECE quantifies this mismatch by binning predictions by confidence and computing the average absolute difference between accuracy and confidence per bin.

Consistent formatting and delimiters (e.g., using triple backticks, XML tags, or clear section headers) help the model parse instructions and context reliably, reducing ambiguity and improving output quality. This is a core best practice in prompt engineering on Vertex AI because it leverages the model's attention mechanisms to focus on distinct prompt segments, leading to more predictable and accurate responses.

Grounding connects the model to verified external data, ensuring responses are based on facts rather than model assumptions. It is a no-retraining approach that directly targets hallucinations by providing context. Other methods like temperature or token limits do not reliably improve factual accuracy.

Input validation and output filtering with safety attributes directly address security vulnerabilities by sanitizing user inputs and filtering model outputs for harmful content. The Vertex AI PaLM API provides safety attribute scores (e.g., toxicity, harassment) that allow developers to programmatically block or flag responses that exceed defined thresholds, reducing the risk of generating insecure code snippets.

Large language models like Gemini are pre-trained on extensive text, including medical content, and can generate coherent, context-aware narratives. They are well-suited for drafting discharge summaries because they understand medical terminology and can adapt to different patient cases. Other model types lack the language generation capabilities needed for this task.

The analyst needs to produce original, varied, and contextually appropriate text for each product. This is achieved through the generative capability of large language models, which predict and generate sequences of words based on the input prompt. Other options describe classification, retrieval, or sentiment analysis—tasks that do not create novel product descriptions.

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.

To enforce safety and privacy rules on generated content, the firm should use Vertex AI's safety filters and configurable thresholds. These settings allow blocking or adjusting responses that violate defined policies, providing a preventive control at the model output level. This directly supports compliance by reducing the chance of leaking customer data or violating internal rules.

Zero-shot prompting lets a pretrained foundation model perform a new task from instructions alone, with no examples or training. Since the team has no labeled data and wants to avoid training, this is the only approach that meets all constraints. Fine-tuning, training from scratch, and rule-based templates either require training or are not generative AI.

Generative AI Studio offers a no-code, console-based environment for prompting and experimenting with foundation models. The marketing team can quickly test multilingual ad copy by entering prompts and adjusting settings, without writing any code. This makes it the fastest and most accessible option for beginners on Google Cloud.

Sampling parameters govern how the model picks the next token. Temperature 0 selects the most probable token deterministically, which is ideal when the same input should map to the same output, as in classification. Other parameters such as token limits and streaming affect output length and delivery, not the randomness of selection, so they cannot deliver the consistency the developer needs.

Safety filters in Vertex AI are specifically designed to detect and block harmful content in both prompts and responses. By configuring these filters, the developer can enforce the company's content policies and reduce the risk of generating or processing unsafe material, ensuring a safer generative AI application.

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.

Vertex AI Studio is the Google Cloud environment built for prompt design and rapid experimentation with generative models. It provides a visual interface where users can enter prompts, compare outputs across models, and adjust sampling parameters. Because the marketing team wants to generate descriptions without coding, this console directly matches their workflow.

Retrieval-augmented generation can be combined with access control lists so that the retrieval step only returns documents the requesting user is permitted to view. This ensures the model's context is limited to authorized information, preventing leakage across departments. Fine-tuning or context expansion without access controls would not solve the security requirement.

A is correct because deploying two models to the same endpoint with traffic splitting is a standard canary deployment strategy. In this configuration, a small percentage of inference requests are routed to the new model while the majority go to the stable model, allowing validation of the new model's performance before full rollout. This is commonly supported by Google Cloud's Vertex AI, where you can deploy multiple models to an endpoint and assign traffic percentages to each model variant (e.g., 90% to the stable model and 10% to the canary model).

Vertex AI Safety Filters are specifically built to detect and filter harmful content in generative AI outputs. For a healthcare application summarizing patient notes, using these filters ensures that inappropriate or harmful content is blocked in real time. Other services like Model Garden, Search, or Pipelines serve different purposes and do not provide content moderation.

Temperature is the parameter that scales the logits before softmax, directly controlling the randomness of the output distribution. Higher values flatten the distribution, making less likely tokens more probable, which increases creativity and variation. For generating diverse blog drafts, the team should increase the temperature setting.

The 429 (Too Many Requests) errors indicate the endpoint is receiving more concurrent requests than its current replica capacity can handle. Increasing MIN_REPLICA_COUNT raises the baseline number of serving replicas, so the endpoint has more capacity to absorb peak traffic without throttling. This directly addresses the throughput bottleneck rather than changing the instance shape or disabling scaling.

The model's training loss and accuracy improve each epoch, but performance on the validation set is poor. This classic symptom indicates overfitting, where the model memorizes the training data (including noise) rather than learning generalizable patterns. In fine-tuning, this often occurs when the model is trained for too many epochs or the dataset is too small relative to model capacity.

Grounding with Google Search is the correct approach because it allows the PaLM 2 API to retrieve real-time, verifiable information from the web, directly reducing factual inaccuracies in generated product descriptions. Unlike parameter adjustments, grounding provides an external knowledge source that the model can cite, ensuring outputs are based on current and accurate data rather than relying solely on its training data.

Gen AI Studio in Vertex AI is the correct tool because it provides a unified interface for discovering, testing, and customizing a wide range of foundation models (e.g., PaLM 2, Gemini, Codey, Imagen) directly from Google Cloud. It allows developers to quickly experiment with different models via a web UI or API without provisioning any infrastructure, making it ideal for rapid prototyping and prompt engineering.

Vertex AI Text Generation (Gemini) is the correct choice because it is a generative AI service specifically designed for natural language understanding and generation tasks, such as summarizing customer support tickets. Gemini models can process long-form text and produce concise, coherent summaries by leveraging transformer-based architectures fine-tuned for instruction-following and text completion. This makes it ideal for extracting key information from support conversations and generating actionable summaries.

Retrieval-augmented generation (RAG) grounds the model's output by retrieving relevant, up-to-date information from a curated database and conditioning the generation on that context. This reduces hallucinations and ensures responses align with approved clinical guidelines. Other options either increase randomness, rely solely on fine-tuning without retrieval, or incorrectly assume smaller models are safer.

Reducing the batch size decreases the memory footprint per training step, allowing the model to fit within the available GPU or TPU memory. Out-of-memory errors during fine-tuning on Vertex AI typically occur when the batch size is too large for the allocated accelerator memory, and lowering it directly resolves the issue without changing the model architecture or hardware.

Prompt engineering allows the model to generate content based on context provided at inference time, so the application can incorporate the latest inventory and brand guidelines without modifying the model. This approach is flexible, cost-effective, and supports frequent updates, unlike fine-tuning or retraining, which are static and resource-intensive.

Grounding with Vertex AI Search connects the Gemini API to a verified, structured data source (e.g., product catalog), enabling the model to retrieve and cite factual specifications rather than relying solely on its training data. This directly addresses the hallucination of wrong dimensions and materials by providing a retrieval-augmented generation (RAG) mechanism that overrides the model's parametric knowledge with authoritative, up-to-date information.

Imagen is the correct choice because it is a text-to-image model in Vertex AI Model Garden specifically designed to generate high-fidelity images from natural language descriptions. Unlike the other options, Imagen uses a diffusion-based architecture to create photorealistic visuals, making it the only option that directly addresses the requirement of generating images from text.

RLHF aligns model behavior with human values by learning from human feedback, while content filters provide a post-generation safety net. Together, they form a robust strategy for reducing harmful outputs. Other options like increasing temperature or limiting output length do not effectively mitigate toxicity and may introduce other issues.

Online prediction with a dedicated endpoint is the correct choice for low-latency, high-throughput serving. It allows the application to send requests and receive immediate responses, with autoscaling to handle varying loads. Batch prediction, Pipelines, and monitoring serve different purposes and cannot fulfill the real-time inference requirement.

Retrieval Augmented Generation (RAG) is the correct approach because it grounds the model's output in verifiable, external data sources. By retrieving product facts from a database in real-time, the system ensures that the generated descriptions are based on accurate information, directly mitigating the risk of hallucination. This method combines the generative power of an LLM with a retrieval step that provides factual context, making it ideal for applications where precision is critical.

Early stopping with a patience of 2 epochs is the correct approach because it directly addresses overfitting by halting training when the validation loss fails to improve for a specified number of epochs. This preserves the model's generalization ability without requiring additional compute or data, which aligns with the limited budget constraint. In Vertex AI, early stopping is a built-in hyperparameter tuning strategy that monitors validation metrics and stops the job to prevent further degradation.

Temperature controls the randomness of token selection. Setting it to zero makes the model choose the highest-probability token each time, yielding deterministic output for the same prompt. Top-p, top-k, and max tokens do not provide the same level of determinism.

Reducing temperature makes sampling more deterministic, so identical prompts tend to yield consistent answers. Grounding the model with retrieved knowledge base passages supplies authoritative context and reduces hallucinated details. Used together, these changes directly target both observed issues: inconsistent responses and content not present in the source material.

Using a system prompt is an effective way to set guardrails for the model's behavior. By explicitly instructing the model to avoid personalized financial advice and to include necessary disclaimers, the firm can help ensure compliance with regulations. This approach is flexible and can be updated as regulations change, unlike fine-tuning which requires retraining.

Prompt engineering is the most suitable approach because it allows the team to specify the desired tone, style, and content directly in the prompt. By providing examples and clear instructions, they can steer the foundation model to produce creative yet on-brand descriptions. This method is fast, cost-effective, and requires no model training.

Hallucination is the term for generative AI producing incorrect or nonsensical information that appears confident. In this scenario, the model's factually incorrect script content is a direct example. Other options like overfitting, bias, or underfitting describe different issues and do not capture the specific problem of generating false but plausible text.

Vertex AI Search and Conversation is the only Google Cloud service listed that provides a managed RAG solution for building conversational assistants grounded in enterprise documents. It securely integrates with Cloud Storage and other data sources, enabling the company to deploy an internal HR assistant without exposing sensitive data externally.

Temperature is the parameter that scales the logits before softmax, directly affecting the probability distribution of the next token. A lower temperature makes the distribution sharper, so the model chooses more likely tokens, resulting in more predictable and less creative outputs. This aligns with the team's goal of reducing exaggerated claims and increasing factual consistency.

Clear instructions and few-shot examples are proven prompt engineering techniques that enhance output quality. They reduce ambiguity and guide the model toward the desired format and content. Other options like high temperature or step-by-step reasoning are not appropriate for precise summarization tasks.

Constrained decoding through a response schema forces the model's tokens to conform to a predefined structure, so the output is guaranteed to be valid JSON with the expected fields. Prompt-level instructions and sampling parameters only influence content probabilistically and cannot guarantee parseable structure. For integration with downstream systems that require strict formatting, the schema-based approach is the only reliable choice.

On-demand serving in Vertex AI charges based on actual consumption and scales automatically, so traffic bursts are handled without provisioning capacity in advance. This satisfies both the cost predictability and the no-infrastructure-management requirements. Dedicated nodes and self-managed clusters bill for reserved capacity regardless of use, and container platforms reintroduce operational responsibilities that the company wants to avoid.

Safety filters and content moderation in Vertex AI allow organizations to define and enforce policies that block or flag toxic, harmful, or inappropriate content generated by the model. This feature uses pre-built and customizable classifiers to evaluate prompts and responses against safety attributes (e.g., hate speech, harassment, sexually explicit content) before returning them to the user, directly addressing the requirement to prevent harmful outputs.

Temperature is the parameter that scales the logits before sampling, directly controlling how random or deterministic the model's token choices are. Increasing temperature makes the model more likely to pick less probable words, yielding more varied and creative captions. The other parameters influence sampling or output length but do not provide the same direct control over variability.

Adding a re-ranking step using a cross-encoder model refines the initial retrieval results by scoring each document against the query more accurately, thus improving relevance. Cross-encoders consider the interaction between query and document, unlike bi-encoders used in vector search, leading to better ranking. The other options are less likely to directly improve relevance.

Retrieval-augmented generation (RAG) with a vector database allows the system to fetch relevant user-specific data, such as past viewing history and video metadata, and use it to condition the model's output. This results in summaries tailored to each user's preferences. Fine-tuning per user is unscalable, and generic prompts or temperature adjustments do not provide true personalization.

Using a smaller model variant directly reduces the number of parameters and computational operations required per inference pass, which lowers latency. In text-to-image models like Imagen or Stable Diffusion, the model size is the primary driver of forward-pass time, so a smaller variant (e.g., fewer layers or reduced latent dimensions) yields faster generation.

Distilled models are smaller, faster versions of larger foundation models, trained to mimic their behavior while requiring fewer computational resources. This directly reduces inference latency because fewer parameters mean faster forward passes through the network, which is critical for real-time chatbot responses.

Few-shot prompting is a powerful, no-retraining method to steer a model's style by providing examples. It allows the model to infer the desired tone and structure from the prompt. Other options either require retraining or adjust parameters that affect randomness, not style adherence.

Vertex AI Search and Conversation and Vertex AI Generative AI Studio are both Google Cloud services that enable grounding in proprietary data and integration into applications. Vertex AI Search and Conversation provides a turnkey solution for grounded search and chat, while Generative AI Studio offers more customization for building and deploying models with retrieval augmentation.

Retrieval-augmented generation (RAG) retrieves relevant information from a trusted source and feeds it to the model, ensuring summaries are based on actual content. This reduces hallucinations by anchoring the model's output to provided facts. Other methods like increasing temperature or fine-tuning do not directly address the need for factual grounding.

Retrieval-Augmented Generation (RAG) is the most direct and effective mitigation for hallucination in document summarization because it forces the LLM to base its output on retrieved chunks of the original document. By requiring citations, the model must reference specific passages, making it verifiable and reducing the likelihood of fabricating content. This grounds the generation in the source material, addressing the root cause of hallucination—lack of factual grounding—rather than relying on post-hoc correction or output tuning.

Retrieval-augmented generation (RAG) allows the model to access and incorporate real-time data from external sources. This ensures responses are based on the most current information, which is critical for market trends. Other options like fine-tuning or increasing temperature do not provide up-to-date data.

Temperature is the sampling parameter that scales the model's predicted token distribution, making outputs more random and creative when increased. Marketing copy benefits from this controlled variability because it produces distinct phrasings from the same feature list. Other parameters such as Top-K, Top-P, and maximum output tokens influence candidate selection or length but do not by themselves introduce the randomness the team wants.

Vertex AI offers managed access to generative models with enterprise controls inside the customer's Google Cloud project. Google Cloud's terms specify that customer data for foundation models is not used to train those models, and encryption in transit and at rest is provided by default. This combination satisfies the governance and no-infrastructure-management requirements.

Increasing the similarity score threshold to 0.85 ensures that only highly relevant chunks are passed to the Gemini 1.5 Pro model, directly reducing the risk of the model generating responses based on irrelevant or low-confidence context. In a RAG pipeline using Vertex AI Search, low-similarity chunks can contain order statuses from different customers or products, leading to hallucinations like incorrect order statuses. Filtering out these less relevant chunks improves the factual grounding of the model's output.

The error shown in the exhibit is an HTTP 403 Forbidden response, which indicates that the server understood the request but refuses to authorize it. In Google Cloud, this is most commonly caused by the user's identity lacking the necessary IAM role or permission to call the specific API or access the resource. Even if the project ID is correct and the network is functional, a missing IAM role (e.g., `aiplatform.user` or `roles/aiplatform.user`) will result in this exact error.

Large language models like Gemini are pre-trained on diverse text and can generate creative, varied outputs from prompts alone. They do not require task-specific examples, making them perfect for producing unique product descriptions at scale. Other model types are designed for different tasks and cannot fulfill the creative text generation requirement.

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

Fine-tuning on a diverse corpus of enforceable contracts teaches the model patterns of valid legal language across contexts, while an external rule-based compliance verifier checks generated clauses against jurisdiction-specific statutes and regulations. This hybrid approach combines generative fluency with deterministic legal validation, which is necessary because LLMs alone cannot guarantee current legal compliance. The verifier provides an auditable, updatable layer that can be revised as laws change.

Few-shot prompting is the most direct way to guide the model to include all features and avoid hallucinations. By showing examples where every feature is incorporated accurately, the model learns the expected pattern. Other parameters like temperature or top-p affect randomness but do not provide the necessary instruction on completeness and fidelity.