AI-102 Azure AI Engineer Associate - Set 2 - Part 1

Correct Answer: During contextual interaction

Containers must connect to Azure to report billing metrics despite local inference processing, ensuring proper usage tracking and compliance.

Option a is incorrect because container registry access is only needed during initial deployment to pull the container image from Microsoft's registry.

Option c is incorrect because GPU availability depends on the specific service being deployed and is not universally required for all containers.

Option d is incorrect because subscription renewal is a general requirement not specific to container deployments versus cloud-based API calls.

Correct Answer: Azure AI Services containers

Containers enable local inference for reduced latency while keeping data on-premise, addressing both performance and security requirements effectively.

Option a is incorrect because Global Standard deployment processes data in Azure cloud regions, not addressing the on-premise data requirement.

Option b is incorrect because Provisioned Managed deployment provides dedicated compute in Azure but doesn't keep data on-premise for processing.

Option d is incorrect because Global Batch deployment is for offline processing and doesn't address latency or on-premise data requirements.

Correct Answer: Storage Accounts repository

Storage Accounts provide cost-effective long-term retention ideal for compliance auditing and archival requirements in production environments.

Option a is incorrect because Log Analytics Workspace is optimized for short-term querying and analysis rather than long-term archival storage.

Option b is incorrect because Event Hubs is designed for real-time streaming to external systems, not long-term storage and archival.

Option d is incorrect because Azure Monitor alerts are for notifications about events, not for storing historical data for auditing.

Correct Answer: Unified multi-model access

Azure AI Studio provides unified access to OpenAI models plus open models like Llama and Mistral, expanding beyond OpenAI-only capabilities.

Option a is incorrect because both studios support similar programming languages through their respective SDKs and REST API interfaces.

Option c is incorrect because deployment regions are determined by Azure infrastructure availability, not the specific studio interface used.

Option d is incorrect because API versioning is managed at the service level, not differentiated between the two studio interfaces.

Correct Answer: Low temperature setting

Low temperature values like 0.1-0.2 produce deterministic, repetitive answers ideal for consistent legal document analysis requirements.

Option a is incorrect because high temperature settings increase randomness and creativity, producing inconsistent outputs unsuitable for legal analysis.

Option c is incorrect because high top-p values increase the token selection pool, leading to more varied rather than consistent responses.

Option d is incorrect because maximum token limit controls response length, not the consistency or determinism of the generated content.

Correct Answer: Variant comparison feature

The variant feature enables creating multiple prompt versions and running them simultaneously to compare results and optimize performance.

Option a is incorrect because batch run testing processes multiple inputs through a single flow configuration, not comparing different prompt variations.

Option b is incorrect because runtime configuration manages compute resources for execution, not comparison of different prompt variations simultaneously.

Option d is incorrect because connection management handles API credentials and endpoints, not the comparison of prompt variations or results.

Correct Answer: Complex domain-specific formatting

Fine-tuning excels when prompt engineering cannot achieve complex formatting or domain-specific tone requirements that need model weight adjustments.

Option a is incorrect because quick prototypes benefit from prompt engineering's immediate results without the time investment of fine-tuning processes.

Option c is incorrect because general knowledge queries work well with base models and don't require the specificity that fine-tuning provides.

Option d is incorrect because simple instructions can typically be handled effectively through well-crafted system messages and few-shot examples.

Correct Answer: Groundedness evaluation metric

Groundedness specifically measures whether generated answers are supported by the provided context, crucial for RAG application reliability.

Option a is incorrect because relevance measures how well the answer addresses the query, not whether it's supported by context.

Option b is incorrect because coherence evaluates how naturally the text reads, not its factual grounding in provided information.

Option c is incorrect because similarity compares generated answers against known correct answers, not their support from context documents.

Correct Answer: Global Batch deployment

Global Batch processes requests offline at reduced costs, perfect for non-real-time batch processing like overnight feedback analysis.

Option b is incorrect because Standard regional deployment is for real-time processing with region-specific pricing, not optimized for batch operations.

Option c is incorrect because Global Standard provides real-time globally available inference at standard pricing, not cost-optimized batch processing.

Option d is incorrect because Provisioned Managed requires paying for dedicated compute regardless of usage, not cost-effective for batch workloads.

Correct Answer: EULA acceptance variable

The EULA environment variable must be set to "accept" to acknowledge Microsoft's licensing terms for container deployment.

Option a is incorrect because API_KEY is for authentication with Azure billing services, not for accepting licensing terms specifically.

Option b is incorrect because BILLING_ENDPOINT provides the URL for billing metrics reporting, not for license agreement acceptance confirmation.

Option d is incorrect because CONTAINER_REGISTRY is not a required environment variable for deployment, only for pulling the container image.

Correct Answer: Log Analytics Workspace

Log Analytics Workspace enables fast querying using Kusto Query Language, ideal for analyzing usage patterns and operational insights.

Option b is incorrect because Storage Accounts are optimized for long-term archival rather than quick querying and real-time analysis capabilities.

Option c is incorrect because Event Hubs streams data to external systems but doesn't provide native querying capabilities like KQL.

Option d is incorrect because Azure Monitor metrics provide numerical data and alerts but not the detailed querying capability of KQL.

Correct Answer: Single key multi-service

Azure AI Services provides one key and endpoint for multiple services, eliminating the need for separate resources per API.

Option a is incorrect because the rebrand primarily focused on resource consolidation rather than fundamental changes to the pricing model structure.

Option b is incorrect because programming language support remained consistent, with SDKs available for the same languages as before rebranding.

Option d is incorrect because regional availability is determined by Azure infrastructure expansion, not the rebranding of the service itself.

Correct Answer: Persona and rules definition

System messages establish the AI's persona and behavioral rules, fundamentally grounding how the model responds to user interactions.

Option a is incorrect because token limits are configured through API parameters, not through the system message content or structure.

Option b is incorrect because while system messages can include format guidance, their primary purpose is establishing persona and behavior rules.

Option d is incorrect because temperature is a separate API parameter that controls randomness, not defined within the system message itself.

Correct Answer: Provisioned Managed deployment

Provisioned Managed deployment allocates dedicated compute resources through Throughput Units, ensuring predictable performance for critical workloads.

Option a is incorrect because Global Batch is for offline processing without guaranteed real-time throughput or dedicated compute resources allocation.

Option b is incorrect because Global Standard shares resources across users globally without dedicated compute guarantees for individual customers.

Option c is incorrect because Standard regional deployment uses shared regional resources without dedicated compute allocation or throughput guarantees.

Correct Answer: flow.dag.yaml file

The flow.dag.yaml file defines the directed acyclic graph structure, specifying how nodes connect and data flows through them.

Option a is incorrect because requirements.txt lists Python dependencies for installation, not the workflow structure or node connections definition.

Option c is incorrect because config.json typically stores application settings but not the specific DAG structure of Prompt Flow workflows.

Option d is incorrect because while connections are important, they're defined within flow.dag.yaml or separately, not in a connections.yml file.

Correct Answer: Parameter Efficient Fine-Tuning

Parameter Efficient Fine-Tuning (PEFT/LoRA) updates only specific layers, reducing computational requirements while maintaining effectiveness for targeted improvements.

Option a is incorrect because full model retraining updates all layers, requiring maximum time and computational resources for the task.

Option b is incorrect because base model selection chooses a starting point but doesn't address the selective layer updating requirement.

Option d is incorrect because embedding adjustment is too specific and doesn't represent the broader PEFT approach for selective layer updates.

Correct Answer: Adjust one parameter only

Best practice recommends adjusting either Temperature or Top P individually, as simultaneous changes can produce unpredictable interaction effects.

Option a is incorrect because increasing both parameters simultaneously can create compounding effects that make behavior difficult to predict or control.

Option c is incorrect because resetting to defaults doesn't address the need to tune parameters for specific use case requirements.

Option d is incorrect because using maximum values would create extremely random outputs, not solving the control and predictability issue.

Correct Answer: Embedding models

Embedding models like text-embedding-3-small convert text into numerical vectors specifically designed for search and similarity operations.

Option a is incorrect because Chat Completion models generate conversational responses, not numerical vector representations for similarity search operations.

Option b is incorrect because DALL-E models generate images from text prompts, not numerical vectors for text similarity comparisons.

Option d is incorrect because Whisper models transcribe speech to text, not create vector embeddings for similarity search functionality.

Correct Answer: Expected correct answer

Ground truth represents the known correct answer used as a benchmark for comparing and scoring the model's generated responses.

Option a is incorrect because source document references are part of context provision, not the ground truth answer for evaluation purposes.

Option c is incorrect because user query intent is implicit in the question formulation, not what ground truth specifically represents.

Option d is incorrect because context relevance score is an evaluation metric output, not the ground truth input for comparison.

Correct Answer: Event Hubs streaming

Event Hubs enables real-time streaming of diagnostic data to external systems including third-party SIEM solutions for integration.

Option a is incorrect because Log Analytics Workspace is for Azure-native analysis, not designed for streaming to external SIEM systems.

Option b is incorrect because Storage Accounts provide static storage for archival, not real-time streaming to external monitoring systems.

Option d is incorrect because Azure Monitor dashboards visualize data within Azure, not integrate with external third-party SIEM solutions.