Business Strategies for Generative AI Solutions practice questions

Limit the system to non-critical administrative tasks only.

Why wrong: This reduces value and does not address liability if the system is used beyond scope.

Use an open-source model to avoid vendor lock-in and reduce costs.

Why wrong: Open-source models may not meet HIPAA compliance and lack support.

Fully automate the system to reduce human error.

Why wrong: Full automation removes human judgment, increasing liability and regulatory risk.

Deploy the model on edge devices to reduce cloud dependency.

Why wrong: Edge devices may lack the compute power for large models and increase maintenance.

Build an on-premises infrastructure to avoid cloud egress fees.

Why wrong: On-premises requires significant capital expenditure and may not scale efficiently.

Provision dedicated GPU instances for consistent performance.

Why wrong: Dedicated GPUs are costly and may be underutilized.

Standardize all descriptions to a neutral tone to avoid cultural issues.

Why wrong: Neutral tone may not engage local audiences effectively.

Use a single global model with a translation layer to convert English descriptions.

Why wrong: Direct translation can lose context and cultural meaning.

The component 'comp-model-eval' does not accept 'eval_dataset' as input

Why wrong: The error is about the parameter not being defined in the pipeline spec, not the component.

The 'project' parameter should be a pipeline input, not a constant

Why wrong: Using a constant for project is valid.

The runtimeConfig parameter values must be strings, not references

Why wrong: The runtimeConfig values are strings; the issue is the missing declaration.

Increase the maximum replicas for the europe-west4 endpoint and reduce the min replicas in other regions

Why wrong: Reducing min replicas in other regions could cause latency issues there during spikes.

Implement Cloud CDN caching for common summaries and reduce the number of regions to two

Why wrong: Cloud CDN is not suitable for dynamic model inference; reducing regions may increase latency for some users.

Configure a global load balancer with a single Vertex AI endpoint and increase max replicas globally

Why wrong: A single endpoint may not reduce latency; still need sufficient replicas in each region.

Implement a two-model architecture: a smaller model for simple queries and a larger model for complex queries, with a router based on query complexity.

Why wrong: While this could improve latency, it increases cost and complexity, and does not directly address hallucinations or compliance.

Switch to a purely fine-tuned model without RAG, and rely on fine-tuning data that excludes PII to ensure compliance.

Why wrong: Without RAG, the model cannot access real-time order data, increasing hallucination risk. Fine-tuning alone cannot guarantee up-to-date or accurate responses.

Use a larger, more powerful LLM with chain-of-thought prompting to improve reasoning and reduce hallucinations, and cache frequent queries to reduce latency.

Why wrong: A larger model increases cost and latency, and chain-of-thought may not resolve hallucinations from missing data. Caching helps latency but not compliance.

Committed use discounts

Why wrong: Requires upfront commitment but still charges variable beyond committed usage.

Free tier

Why wrong: Too limited for production use.

Pay-as-you-go

Why wrong: Uses variable pay-per-use billing, not predictable.

Use Vertex AI Endpoint with monitoring

Why wrong: Monitors performance, not cost allocation.

Implement a custom cost allocation using labels

Why wrong: Labels help but are less structured than separate projects.

Use Cloud Billing budgets and alerts per model

Why wrong: Budgets track overall cost, not per model usage.

BLEU score

Why wrong: Measures similarity to reference, not security.

Human evaluation

Why wrong: Subjective and not scalable for vulnerability detection.

Perplexity

Why wrong: Measures model certainty, not code correctness or security.

Minimizing latency at any cost

Why wrong: Minimizing latency at any cost can be too expensive and may not address the primary business consideration of data privacy and compliance.

Using open-source models only

Why wrong: Using open-source models only may not meet all requirements, such as compliance and support, and doesn't address the primary business consideration of data privacy.

Choosing the most complex model

Why wrong: Choosing the most complex model introduces unnecessary computational overhead and latency, which directly conflicts with the need to process a large volume of customer interactions in real time. A lighter, fine-tuned model would provide faster inference and lower operational cost. This option is tempting because complex models excel at nuanced, open-ended tasks, such as legal document analysis, where depth of reasoning outweighs throughput requirements.

Remove the binding

Why wrong: Removing the binding removes all access, which is too restrictive.

Add more roles

Why wrong: Adding roles does not restrict; it grants additional permissions.

Use organizational policies

Why wrong: Organizational policies set constraints but are not as granular as IAM conditions for specific actions.

No effective permissions because the role is incorrect.

Why wrong: The role is valid and grants specific permissions.

Ability to deploy and manage models.

Why wrong: Deploying models requires roles/aiplatform.deployer or higher.

Full control over all Vertex AI resources.

Why wrong: That requires roles/aiplatform.admin, not user.

Fine-tune Gemini with their product data using Vertex AI Generative AI Studio.

Why wrong: Fine-tuning Gemini requires substantial labelled data and ML expertise to avoid catastrophic forgetting or overfitting, yet the scenario specifies a small dataset and limited ML expertise. This approach is tempting because Vertex AI Generative AI Studio is designed for custom model adaptation, and it would be correct if the company had a large, high-quality labelled dataset and in-house ML engineers to manage the fine-tuning pipeline.

Build a custom transformer model using TensorFlow on Vertex AI Workbench.

Why wrong: Building a custom transformer model from scratch is complex and data-intensive, not ideal for their constraints.

Use BigQuery ML to train a classification model on customer queries.

Why wrong: BigQuery ML is for structured data classification, not for chatbot or recommendation generation; it lacks conversational capabilities.

Upgrade to GPU-accelerated machines for all replicas.

Why wrong: Incorrect. Upgrading to GPU-accelerated machines increases cost significantly and does not directly solve the cold start issue. While GPUs may reduce inference latency per request, they do not prevent the initial delay when an instance is first created.

Implement client-side request batching to reduce the number of inference calls.

Why wrong: Incorrect. Client-side request batching reduces the number of inference calls but does not address cold start latency. Batching may help with throughput, but individual requests still face cold start delays if no instances are ready.

Use prewarmed containers by setting an idle timeout to keep instances alive.

Why wrong: Incorrect. Prewarmed containers via idle timeout is not configurable in Vertex AI Prediction; the idle timeout is fixed. Additionally, this approach would keep instances alive only briefly and not guarantee availability for infrequent traffic patterns, increasing cost without fully eliminating cold starts.

Enable streaming responses to get real-time answers.

Why wrong: Streaming responses affect how answers are delivered (incrementally vs. all at once), not the source of knowledge used to generate them. It does not constrain the model to company documents.

Fine-tune the model on the company's documents.

Why wrong: Fine-tuning adapts the model's behavior over time using training data, but the question asks about grounding—ensuring answers come from a specific datastore at inference time. Fine-tuning is not the essential configuration for this purpose.

Set the model's temperature to 0 to make responses deterministic.

Why wrong: Temperature controls output randomness and creativity, not the knowledge source. Setting it to zero makes responses deterministic but does not guarantee they come from the company's documents.

Use Video Intelligence API to generate video captions, then feed captions to a text model.

Why wrong: Incorrect. Using Video Intelligence API to generate video captions introduces an additional API and loses visual context from captions, making it less scalable and cost-effective than directly processing with a multimodal model.

Convert all inputs to text descriptions and use a text-only model.

Why wrong: Incorrect. Converting all inputs to text descriptions loses multi-modal information (e.g., visual cues, tone, context), reducing summary accuracy and failing to leverage the full capabilities of a multimodal model.

Deploy a single Vertex AI endpoint with a model that can ingest multi-modal data directly.

Why wrong: Incorrect. Deploying a single Vertex AI endpoint that ingests raw video directly would lead to prohibitive token costs and latency, especially for variable-length videos up to 10 minutes, making it impractical compared to a pipeline with preprocessing and caching.

Use Gemini 1.5 Pro for the highest accuracy.

Why wrong: Gemini 1.5 Pro is the most capable model for accuracy but has higher latency and cost, which contradicts the startup's priorities of low latency and cost.

Use PaLM 2 Bison, as it is the most economical.

Why wrong: PaLM 2 Bison is an older model; while economical, it is not optimized for the high-throughput, low-latency summarization task compared to Gemini 1.5 Flash.

Use Vertex AI Text Embeddings, since embeddings can generate summaries.

Why wrong: Vertex AI Text Embeddings are designed to generate vector representations of text, not summaries. They are not suitable for this task.

Deploy a custom model on Cloud Run and write a Cloud Function to mask PII.

Why wrong: Incorrect. Deploying a custom model on Cloud Run and using a Cloud Function to mask PII requires custom development and does not provide integrated logging and compliance features. It lacks the automated PII masking and audit trails that managed services like Vertex AI Agent Builder offer.

Use Vertex AI Prediction with a custom container that masks PII before inference.

Why wrong: Incorrect. Using Vertex AI Prediction with a custom container that masks PII before inference still requires custom development and does not provide built-in logging or compliance features. It also does not integrate with Data Governance for automated DLP.

Use the Gemini API directly with a custom logging solution in Cloud Logging.

Why wrong: Incorrect. Using the Gemini API directly with custom logging in Cloud Logging requires manual implementation of PII masking and logging, and it does not leverage Vertex AI's managed data governance capabilities, increasing complexity and risk for compliance.

Use AutoML to train a new model directly from the dataset without fine-tuning.

Why wrong: AutoML is designed for training models from scratch with automated hyperparameter tuning and architecture search, not for fine-tuning an existing large language model. It would not leverage the proprietary dataset or the base model, resulting in a fundamentally different approach that does not meet the requirement of fine-tuning the existing LLM.

Deploy the existing pipeline on a Google Kubernetes Engine cluster and use Google Cloud Filestore for shared storage.

Why wrong: Deploying the existing pipeline on Google Kubernetes Engine with Filestore for shared storage requires significant operational overhead for cluster management and storage scaling. While possible, it does not provide the incremental sync capability of Vertex AI Managed Datasets, leading to higher data transfer costs and potential downtime during full data migrations. This is not the most optimized solution for minimizing downtime and cost.

Use Cloud Storage Transfer Service to move all data to Cloud Storage, then set up a Vertex AI custom training job to run the fine-tuning.

Why wrong: Cloud Storage Transfer Service is intended for one-time or scheduled batch transfers of data to Cloud Storage. It would require a full initial data transfer, incurring high costs and potential downtime, and does not natively support incremental syncing of new data for weekly fine-tuning. This approach lacks the integration with BigQuery and the incremental capabilities that Vertex AI Managed Datasets provide.

Set the model temperature to 0 to eliminate creativity and reduce bias.

Why wrong: Setting temperature to 0 reduces randomness but does not eliminate bias or ensure fairness. This strategy does not address dynamic, context-dependent toxic outputs or regulatory compliance.

Fine-tune the model exclusively on compliant financial documents.

Why wrong: Fine-tuning on compliant financial documents only limits the training data but does not guarantee avoidance of toxic outputs or compliance with regulations in all contexts. It does not address input or output filtering.

Disable request logging to avoid storing sensitive data.

Why wrong: Disabling request logging would violate GDPR and other auditing requirements, as logging is necessary for monitoring, compliance, and incident response. It does not enhance safety or fairness.