Google Cloud's Generative AI Offerings practice questions

Configure 'Slot filling' to collect more info

Why wrong: Slot filling is for gathering parameters, not human handoff.

Implement a 'Confirmation' prompt for the user

Why wrong: Confirmation is for verifying actions, not handing off.

Use a 'Fallback' intent to route to a human

Why wrong: Fallback intent is for unrecognized inputs, not specifically for human handoff.

Move the vector search index to a multi-regional Cloud Storage bucket (e.g., 'us') to reduce latency for index updates.

Why wrong: Moving the index to a multi-regional bucket (e.g., 'us') does not reduce latency for Australian users because the bucket is still located in the US. Australia remains far from the data, so response times will still spike.

Deploy a separate vector search index endpoint for each region with its own index copy stored in a regional bucket in that region.

Why wrong: Deploying separate vector search indices with regional copies would introduce data inconsistencies across regions for the global product documentation, failing the implicit need for a unified RAG source. The existing issue stems from the single `us-central1` data source causing high latency for Australian users, despite local endpoints. This option is tempting as it addresses data locality, and would be appropriate if each region required entirely distinct, independent product documentation indices, where cross-regional consistency was not a requirement.

Increase the number of replicas for the vector search index in all regions to improve throughput and reduce latency.

Why wrong: Increasing replicas improves throughput but does not address the geographic distance between the serving region and the storage bucket. The latency spike is due to network distance, not insufficient replicas.

Enable grounding with security scanning tools

Why wrong: Grounding with scanning tools is not a built-in feature.

Use the Vertex AI Codey API with safety settings

Why wrong: Safety settings in Codey API are for content safety, not code security.

Use a custom safety attribute filter

Why wrong: Safety filters address harmful content, not code vulnerabilities.

Set safety_threshold high for all categories

Why wrong: Incorrect. A high threshold (BLOCK_ONLY_HIGH) only blocks content with high harm probability, allowing medium- and low-probability harmful content through, which is insufficient for strict safety.

Use grounding with a safe knowledge base

Why wrong: Grounding provides context but does not block harmful generation.

Fine-tune the model with safe examples

Why wrong: Fine-tuning helps but does not guarantee safety; safety filters are still needed.

Increase the max output tokens

Why wrong: Increasing output tokens may increase cost.

Use grounding with Google Search

Why wrong: Grounding with Google Search incurs additional costs.

Use a larger model

Why wrong: Larger models cost more per token.

BigQuery

Why wrong: BigQuery is for analytics, not required for the agent.

Vertex AI Prediction

Why wrong: Prediction is for serving models, not building conversational agents.

Cloud Storage

Why wrong: Cloud Storage is for storing data, but not essential for the agent's conversational logic.

Number of model versions

Why wrong: Number of versions is not a key selection criterion.

The color of the model card

Why wrong: Color is irrelevant.

The model's artifact URI is wrong

Why wrong: Artifact URI issues would cause a different error during deployment.

The model wasn't uploaded correctly

Why wrong: The error indicates the model exists but lacks a serving container.

The model is in a different project

Why wrong: Cross-project listing would require --project flag, not give this error.

The user needs more permissions

Why wrong: The user has admin role; the issue is the agent's service account.

The agent needs a bigger quota

Why wrong: Quota issues would give a quota error, not an access error.

The data store is not in the same region

Why wrong: Data store regions must match, but that would cause a different error.

The endpoint returns a 4xx status

Why wrong: A 4xx error would still trigger the action but return an error; the action would be attempted.

The endpoint is not HTTPS

Why wrong: The endpoint uses HTTPS, which is required.

The agent is not enabled

Why wrong: The agent exists in the YAML; if disabled, a different error would occur.

Increase the number of training steps

Why wrong: More steps increase training time and cost.

Increase the batch size

Why wrong: Larger batch sizes can slow down training and may degrade model quality if not adjusted properly.

Use a higher learning rate

Why wrong: A higher learning rate may cause instability and poor convergence.

Google Cloud documentation

Why wrong: Documentation may not include the most current policy.

Google's AI Principles

Why wrong: AI Principles are high-level, not a detailed policy.

The Gemini Terms of Service

Why wrong: Terms of Service for a specific product may not cover all services.

Developer preference

Why wrong: Developer preference is not a primary enterprise consideration.

Model size

Why wrong: While relevant, model size is less critical than latency and data residency.

Training time

Why wrong: Training time is not relevant for model deployment.

Disable content filtering

Why wrong: Disabling filtering increases risk of harmful outputs.

Use only Google's foundation models

Why wrong: Even Google's models need responsible practices.

Color space

Why wrong: Color space is for images.

Audio sample rate

Why wrong: Audio sample rate is for audio processing.

Image size

Why wrong: Image size is for image generation.

Deploy the agent to a Vertex AI endpoint

Why wrong: Endpoint deployment is for serving the model, not connecting to external data.

Fine-tune a Gemini model on the knowledge base

Why wrong: Fine-tuning would adapt the model but is less efficient for dynamic retrieval; grounding is designed for this use case.

Configure a safety filter to block irrelevant queries

Why wrong: Safety filters block harmful content, they do not retrieve information.

Use A/B testing between versions and manually select the best performer

Why wrong: A/B testing requires manual intervention and does not automatically promote based on thresholds.

Manually track version IDs and deploy the latest version

Why wrong: Manual tracking is not automated and can lead to deploying untested versions.

Deploy all versions to a single endpoint and route traffic manually

Why wrong: Manual routing is not automated and increases risk of deploying poor versions.

Increase the temperature parameter to make the model more creative

Why wrong: Higher temperature increases randomness, likely increasing hallucinations.

Reduce the maximum output tokens to force concise answers

Why wrong: Reducing tokens may truncate useful information and does not address factual accuracy.

Adjust safety settings to filter uncertain responses

Why wrong: Safety settings filter based on content categories, not correctness.

Vertex AI Search with grounding

Why wrong: Search is for retrieval, not for customizing generation style.

Prompt design with the Gemini API directly

Why wrong: Prompt design is a method, but for deep customization, tuning is more effective.

Vertex AI Model Evaluation

Why wrong: Model Evaluation is for testing performance, not for customization.

Vertex AI Model Registry

Why wrong: Model Registry manages versions, not explainability.

Vertex AI Safety Settings

Why wrong: Safety Settings filter content, but do not explain outputs.

Vertex AI AutoML

Why wrong: AutoML automates model training but does not natively provide explanations.