Generative AI Leader Google Cloud's Generative AI Offerings — Timed 40 Questions

Question 1

A financial services firm uses a fine-tuned Gemini model in Vertex AI for regulatory compliance checks. They notice that token usage is high, increasing costs. They want to reduce costs without sacrificing accuracy. Which approach should they take?

40 minute time limit — choose an answer to begin.

Accepted Answer

Set max output tokens to a lower value and use more precise prompts. Reducing max output tokens directly lowers the number of tokens generated per request, which is the primary cost driver in pay-per-token models like Gemini. Using more precise prompts further reduces token waste by guiding the model to produce concise, relevant outputs without sacrificing accuracy, as compliance checks often require specific, structured responses rather than verbose explanations.

Answer

Switch to a smaller base model like PaLM 2 Bison

Answer

Enable context caching to reuse previous responses

Answer

Reduce temperature to 0.0

Question 2

What is the most likely cause of the error?

Accepted Answer

The predict schema must be stored in the same bucket as the model artifacts and referenced without the full gs:// URI. The error occurs because the Vertex AI Predict schema must be stored in the same Cloud Storage bucket as the model artifacts, and when referenced in the model upload request, it should use a relative path (without the full `gs://` URI). Using the full URI causes a parsing failure, as Vertex AI expects the schema to be co-located with the model artifacts for validation and deployment.

Answer

The display name contains a hyphen which is not allowed

Answer

The container image URI is incorrect

Answer

The region us-central1 does not support TensorFlow models

Question 3

Why is the model responding in English despite the prompt asking for French translation?

Accepted Answer

The system instruction to translate to French was not set; the user prompt alone is not sufficient. In Google Cloud's Vertex AI and Generative AI offerings, the system instruction is a separate, persistent directive that sets the model's behavior, such as language output. The user prompt alone, even if it asks for a French translation, is not sufficient to override the default language of the model; the system instruction must explicitly specify the target language. Without this instruction, the model defaults to its training language (typically English), regardless of the user's request.

Answer

The model endpoint is configured for English only

Answer

The temperature is too high, causing random outputs

Answer

The maxOutputTokens is too low to complete the translation

Question 4

A financial services firm needs to deploy a large language model (LLM) for analyzing sensitive client documents. They require the model to run within their Virtual Private Cloud (VPC) with no internet access and must comply with data residency regulations. Which Google Cloud generative AI offering should they use?

Accepted Answer

Vertex AI Model Garden with private endpoints and VPC Service Controls. Vertex AI Model Garden with private endpoints and VPC Service Controls allows the LLM to be deployed entirely within the customer's VPC, with no internet egress, and enforces data residency by restricting data movement to the configured VPC boundary. Private endpoints use Private Service Connect to route inference traffic through internal IPs, while VPC Service Controls prevent data exfiltration and ensure compliance with residency regulations.

Answer

Vertex AI Search

Answer

Cloud Run

Answer

Vertex AI Workbench

Question 5

A company is using Vertex AI for multimodal generative AI to analyze images and text. They need to ensure that the model's outputs are auditable and can be traced back to the input data. Which feature should they enable?

Accepted Answer

Vertex AI Model Monitoring with Explainable AI. Vertex AI Model Monitoring with Explainable AI provides feature attributions that map model predictions back to specific input features (e.g., pixels in images or tokens in text). This creates an auditable trail by quantifying how each input contributed to the output, enabling traceability for compliance and debugging. The other options lack the direct input-to-output attribution required for auditability.

Answer

Vertex AI Feature Store

Answer

Vertex AI Experiments

Answer

Cloud Logging

Question 6

A team deployed a custom generative AI model using KServe on Google Kubernetes Engine (GKE) with the above configuration. They notice that the model is taking longer than expected to respond. What is the most likely cause?

Accepted Answer

The model requires more than 1 GPU for acceptable performance. The configuration specifies 1 GPU, but the model requires more than 1 GPU for acceptable performance. KServe on GKE allocates GPU resources based on the `limits` field; if the model's inference workload exceeds the memory bandwidth or compute capacity of a single GPU, latency increases due to queuing and serialization. This is the most likely cause of the slow response time, as GPU-bound models are sensitive to under-provisioning.

Answer

The CPU resource limits are too low

Answer

The model is crashing due to insufficient memory

Answer

The container image is too large and takes time to pull

Question 7

You are a machine learning engineer at a healthcare startup. Your team has developed a generative AI model that summarizes patient medical records. The model is deployed on Vertex AI Endpoints using a custom container. You have configured the endpoint with a single n1-standard-4 machine (4 vCPUs, 15 GB memory) without accelerators. The model uses a small transformer architecture. During load testing with 50 concurrent requests, you observe that the average latency is 8 seconds, which exceeds the requirement of 2 seconds. Additionally, some requests time out after 10 seconds. You suspect the CPU is the bottleneck. You also notice that the model inference code uses TensorFlow but is not optimized for inference. Which action should you take to reduce latency?

Accepted Answer

Deploy the model on a machine with a GPU and use TensorRT for inference optimization.. The CPU is identified as the bottleneck, and deploying on a GPU with TensorRT optimization directly addresses this by accelerating the TensorFlow inference. TensorRT optimizes the model graph and fuses layers, significantly reducing latency for transformer-based models, which is essential to meet the 2-second requirement.

Answer

Reduce the model size by pruning and quantization, then redeploy.

Answer

Increase the request timeout to 30 seconds to accommodate the latency.

Answer

Enable autoscaling to handle the load with multiple replicas.

Question 8

A data scientist is using Vertex AI Model-as-a-Service (MaaS) to deploy a fine-tuned open-source model. They notice high latency during inference. What is the most likely cause?

Accepted Answer

The endpoint is set to autoscaling with a low minimum node count. High latency during inference on Vertex AI MaaS is most often caused by the endpoint scaling configuration. When autoscaling is enabled with a low minimum node count, the system may need to provision additional nodes to handle the request load, which introduces cold-start latency. This is especially pronounced for fine-tuned open-source models, which can be large and take time to load onto new nodes.

Answer

The model is too large for the hardware

Answer

The model is not quantized

Answer

The region is incorrect

Question 9

A financial services company wants to use Vertex AI Grounding with enterprise data to power a regulatory compliance chatbot. They have strict data residency requirements: data must remain in the EU. What should they do?

Accepted Answer

Enable Data Residency by selecting a EU region during data store creation. Vertex AI Grounding with enterprise data requires a data store, and when creating that data store in Vertex AI Search, you can select a EU region (e.g., europe-west1) to enforce data residency. This ensures all indexed enterprise data and grounding operations remain within the EU, satisfying strict regulatory requirements. The data store region determines where the data is stored and processed, independent of the Vertex AI endpoint region used for model inference.

Answer

Use a VPN to the US region

Answer

Convert data to private tokens

Answer

Use a Vertex AI endpoint in a European region

Question 10

You are using Vertex AI Model Garden to deploy a Llama model. Which deployment option provides the best latency for real-time inference?

Accepted Answer

Deploy to Vertex AI Endpoint with a fixed number of replicas. Deploying a Llama model to a Vertex AI Endpoint with a fixed number of replicas ensures that compute resources are pre-allocated and always warm, minimizing cold-start latency and providing consistent, low-latency responses for real-time inference. This approach uses a dedicated endpoint with persistent instances, which is optimized for sub-second response times required by interactive applications.

Answer

Use Batch Prediction

Answer

Deploy to a Compute Engine VM

Answer

Use MaaS (Model-as-a-Service) with autoscaling

Question 11

A team is deploying a real-time chat application using Gemini. They need to ensure the model does not generate harmful content. Which safety filter configuration should they use?

Accepted Answer

Implement custom safety attribute filters with low thresholds. Setting safety thresholds to low (e.g., BLOCK_LOW_AND_ABOVE) for all categories blocks even low-probability harmful content, ensuring strict safety in a real-time chat application. Option A is incorrect because a high threshold (e.g., BLOCK_ONLY_HIGH) only blocks content with high probability of being harmful, allowing potentially harmful content through. Option B is incorrect because grounding with a safe knowledge base limits the knowledge source but does not filter the model's generated responses. Option D is incorrect because fine-tuning with safe examples reduces but does not eliminate the need for active safety filtering during inference.

Answer

Set safety_threshold high for all categories

Answer

Use grounding with a safe knowledge base

Answer

Fine-tune the model with safe examples

Question 12

Which TWO components are essential for building a multi-turn conversational agent using Vertex AI Agent Builder? (Choose two.)

Accepted Answer

Dialogflow CX. Dialogflow CX (C) is essential because it provides the conversational orchestration layer for multi-turn dialogue, including flows, pages, intents, and state/parameter management that track context across turns. Agent Builder Agent (D) is essential because it is the core agent construct in Vertex AI Agent Builder that defines the agent's instructions, tools, and behavior for handling multi-turn conversations. Together, Dialogflow CX and the Agent Builder Agent supply the dialogue state management and agent definition needed for a multi-turn conversational agent. BigQuery (A) is a data warehouse for analytics, not a required conversational component; Vertex AI Prediction (B) serves deployed ML models but is not required to build the conversational agent; and Cloud Storage (E) is object storage for files, not a required part of the agent's conversational logic.

Answer

BigQuery

Answer

Vertex AI Prediction

Answer

Cloud Storage

Question 13

Refer to the exhibit. You ran the gcloud command to list a model, but received this error. What is the most likely issue?

Accepted Answer

The model is missing a serving container. The error indicates that the model cannot be deployed because it lacks a serving container specification. In Vertex AI, models must have either a pre-built or custom serving container defined in the model resource to handle prediction requests. Without this, the gcloud command cannot list or deploy the model, making option C correct.

Answer

The model's artifact URI is wrong

Answer

The model wasn't uploaded correctly

Answer

The model is in a different project

Question 14

Refer to the exhibit. A developer has defined a dynamic action in the Vertex AI Agent Builder agent YAML. The agent is not triggering the action. What is the most likely issue?

Accepted Answer

The action name is misspelled. The most common reason a dynamic action fails to trigger in Vertex AI Agent Builder is a mismatch between the action name defined in the agent YAML and the name referenced in the agent's configuration or invocation logic. The YAML parser is case-sensitive and requires exact string matching; even a minor typo or capitalization error will cause the agent to ignore the action entirely, as it cannot resolve the reference to the defined action block.

Answer

The endpoint returns a 4xx status

Answer

The endpoint is not HTTPS

Answer

The agent is not enabled

Question 15

A company is deploying a chatbot that must ensure customer data remains within the European Union. Which approach should they take?

Accepted Answer

Use the Gemini API with a regional endpoint in europe-west4. Using the Gemini API with a regional endpoint in europe-west4 ensures that all data processing and storage occurs within the European Union. Regional endpoints restrict data residency to the specified region, complying with GDPR and other EU data sovereignty requirements.

Answer

Use Vertex AI Agent Builder with global endpoint

Answer

Use Vertex AI with a multi-region endpoint

Answer

Deploy a custom model on GKE in a specific region

Question 16

During a load test, a Vertex AI endpoint serving a large language model experiences high latency and increased error rates. The endpoint is configured with autoscaling. What is the most likely cause?

Accepted Answer

The autoscaling metric is based on CPU utilization but the model is GPU-bound. When a model is GPU-bound, CPU utilization remains low even under heavy inference load, so autoscaling based on CPU metrics fails to trigger additional replicas. This leads to queued requests, increased latency, and eventual error rates as the existing GPU instances become saturated. Vertex AI endpoints default to CPU-based autoscaling unless explicitly configured with GPU metrics like 'gpu_utilization' or custom metrics.

Answer

There is a network bottleneck

Answer

The model size is too large for the machine type

Answer

The endpoint is using a global load balancer

Question 17

A data scientist is using Vertex AI Model Registry to manage multiple versions of a custom text classification model. They need to ensure that only the version that passes all evaluation metrics can be deployed to a Vertex AI Endpoint for online predictions. What deployment strategy should they use?

Accepted Answer

Set up continuous evaluation with model monitoring to auto-promote versions that meet thresholds. Vertex AI Model Registry supports continuous evaluation pipelines that automatically promote a model version to a 'default' or 'deployed' status only when it meets predefined evaluation thresholds (e.g., accuracy, F1 score). This ensures that only the version passing all metrics is eligible for deployment to a Vertex AI Endpoint, enabling automated, gated rollouts without manual intervention.

Answer

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

Answer

Manually track version IDs and deploy the latest version

Answer

Deploy all versions to a single endpoint and route traffic manually

Question 18

Which TWO actions can help reduce latency for an online prediction endpoint served by a large language model on Vertex AI? (Select TWO.)

Accepted Answer

Enable response caching for frequent similar queries. Option A is correct because enabling response caching lets the endpoint return previously computed answers for frequent similar queries immediately, bypassing the model's forward-pass computation and thereby cutting inference latency. Option D is correct because a smaller distilled version of the model has fewer parameters and requires less compute per token, so it generates predictions faster than the large teacher model while retaining much of its quality. Option B is not correct because raising the max input token limit only allows longer prompts; it does not reduce latency and can actually increase it by processing more tokens. Option C is not correct because automatic scaling adds replicas to handle higher request volume or throughput, not to lower the per-request latency of a single prediction. Option E is not correct because preemptible (Spot) instances can be reclaimed at any time, which hurts availability and does not improve inference latency.

Answer

Increase the max input token limit to capture more context

Answer

Use automatic scaling to add more replicas

Answer

Set the model to preemptible instances

Question 19

Which THREE capabilities does Vertex AI Agent Builder provide out of the box? (Select THREE.)

Accepted Answer

Automatic escalation to human agents when confidence is low. Vertex AI Agent Builder provides out-of-the-box capabilities including automatic escalation to human agents when confidence is low (A), which is a built-in feature of its conversational agents to hand off to live representatives. It also supports grounding with enterprise data sources such as Cloud Storage and BigQuery (D), allowing agents to retrieve and ground responses in an organization's own data. Additionally, it offers pre-built customer service and sales agents (E), which are ready-to-use templates for common business scenarios. Fine-tuning of foundation models using custom datasets (B) is a capability of Vertex AI's model tuning services, not a specific out-of-the-box feature of Agent Builder. Enterprise search over internal documents (C) is provided by Vertex AI Search (formerly Enterprise Search), which is a related but separate product from Agent Builder.

Answer

Fine-tuning of foundation models using custom datasets

Answer

Enterprise search over internal documents

Question 20

A machine learning engineer submits the above batch prediction job for a large language model. The job is expected to process 100,000 instances. The job takes much longer than expected. Which change would most likely reduce the execution time?

Accepted Answer

Increase maxReplicaCount to 10. Increasing maxReplicaCount to 10 allows Vertex AI Batch Prediction to scale out to more worker replicas, processing the 100,000 instances in parallel. The default maxReplicaCount is often 1 or a low number, which forces sequential or limited parallel processing, causing long execution times. By raising this limit, the job can leverage horizontal scaling to distribute the workload across multiple machines, significantly reducing wall-clock time.

Answer

Increase startingReplicaCount to 10 without changing maxReplicaCount

Answer

Increase the machine type to n1-standard-16

Answer

Decrease the batch size to 1