AI-300 Practice Test 5 — 25 Questions

Question 1

You want to automate the deployment of a model using GitHub Actions. Which file format is standard for defining the Azure ML CLI v2 deployment configuration?

Accepted Answer

YAML. Azure ML CLI v2 uses YAML files to define configurations for jobs, endpoints, and deployments.

Answer

JSON

Answer

XML

Answer

Bicep

Question 2

What is the purpose of the Azure ML 'Model Registry'?

Accepted Answer

Storing model version history. The Registry manages versioning and metadata of trained models.

Answer

Managing network access

Answer

Calculating accuracy metrics

Answer

Executing training jobs

Question 3

You are experiencing throttling on your AI endpoint. Which TWO steps should you take?

Accepted Answer

Check current usage against quota limits.. Option A is correct because throttling on an AI endpoint typically occurs when request rate or token consumption exceeds the provisioned quota, so you must first check current usage against quota limits (e.g., requests per minute, tokens per minute) to confirm which limit is being hit. Option D is correct because, once you confirm the workload legitimately exceeds the assigned quota, the standard remediation is to request a quota increase for the model through the service's quota management process. Option B is wrong because disabling security features does not affect throttling and would expose the endpoint to risk. Option C is wrong because swapping to a faster but less capable model changes capability, not the quota or throttling behavior, and is not a proper fix. Option E is wrong because deleting existing deployments would remove capacity and disrupt service rather than resolve throttling.

Answer

Disable all security features.

Answer

Change the model version to one that is faster but less smart.

Answer

Delete all existing deployments.

Question 4

Which TWO of the following storage types can be registered as an Azure ML Datastore?

Accepted Answer

Azure Blob Storage. Azure Blob Storage (B) is correct because it is one of the core Azure storage services natively supported by Azure Machine Learning for registering as a datastore, allowing ML workspaces to reference blob containers for datasets, training data, and model artifacts via the azureml-core SDK or studio. Azure Data Lake Storage Gen2 (C) is also correct because it is a supported datastore type in Azure ML, providing hierarchical namespace access over blob storage for big-data analytics and ML workloads. Azure App Service (A) is a PaaS web hosting platform, not a storage service, so it cannot be registered as a datastore. Microsoft Entra ID (D) is an identity and access management service, not a storage backend. Azure Bot Service (E) is a conversational AI service, not a storage service, and therefore cannot be registered as an Azure ML datastore.

Answer

Azure App Service

Answer

Azure Active Directory

Answer

Azure Bot Service

Question 5

You want to measure 'Relevance' in a RAG application. The relevance evaluator detects how well the response answers the user query. If the model provides a factually correct answer that does not address the prompt, which metric will capture this failure?

Accepted Answer

Relevance. Relevance specifically measures if the response directly addresses the user's intent.

Answer

Fluency

Answer

Coherence

Answer

Groundedness

Question 6

A user wants to restrict the model's output to valid JSON format. What is the most effective way to ensure this?

Accepted Answer

Enable JSON mode/Response Format in the API configuration.. Enabling JSON mode (Response Format) in the API configuration is the most effective way to constrain model output to valid JSON because the model is instructed and, in supported models, constrained to emit syntactically valid JSON. This is a first-class API feature (response_format: { type: 'json_object' }) that enforces structure at generation time. Prompt instructions alone are probabilistic, and post-processing only fixes output after the fact.

Answer

Increase the temperature.

Answer

Add 'output must be JSON' to the prompt.

Answer

Post-process the output with a regex script.

Question 7

You are managing model lifecycle security. Which THREE actions are recommended to secure your ML models?

Accepted Answer

Store secrets in Azure Key Vault.. Option C is correct because storing secrets (storage account keys, connection strings, service principal credentials) in Azure Key Vault keeps them out of notebooks, scripts, and source control, and lets the workspace retrieve them via its managed identity rather than hard-coded values. Option D is correct because Azure RBAC enforces least-privilege access to the workspace, compute, datastores, and models, ensuring only authorized identities can read, train, deploy, or manage ML assets. Option E is correct because private endpoints place the workspace and its associated resources (such as storage, Key Vault, and container registry) on a private IP within your VNet, removing public internet exposure and blocking exfiltration. Option A is not recommended because allowing all traffic in NSGs defeats network segmentation and exposes the workspace to unauthorized access. Option B is not recommended because making datastores public exposes training data and model artifacts to anonymous or unauthenticated access, violating data-protection requirements.

Answer

Allow all traffic in NSGs.

Answer

Make all Datastores public.

Question 8

Your team is using Semantic Kernel for agent orchestration. You notice frequent transient errors when the agent interacts with multiple LLM services. How should you implement infrastructure-level resiliency?

Accepted Answer

Use an 'Azure API Management' policy to implement a retry pattern.. Using an Azure API Management (APIM) policy to implement a retry pattern is the correct approach for infrastructure-level resiliency. APIM acts as a gateway that can intercept requests to backend LLM services and apply policies such as retry, circuit breaker, and rate limiting. This offloads resiliency logic from the application code and provides a centralized, configurable way to handle transient errors across multiple LLM services.

Answer

Disable the 'Streaming' flag in the agent request.

Answer

Switch the model to a higher tier instance type.

Answer

Increase the 'Timeout' parameter in the model deployment manifest.

Question 9

You are configuring a chat application. What is the benefit of enabling streaming in the Azure OpenAI API?

Accepted Answer

Reduces time-to-first-token (TTFT) perception for the end-user.. Streaming improves perceived latency by delivering tokens as they are generated.

Answer

Increases the model's accuracy.

Answer

Removes the need for a system prompt.

Answer

Decreases the total number of tokens consumed.

Question 10

You are deploying a large model. During the deployment, you encounter a 'Resource Not Available' error. What is the most likely cause?

Accepted Answer

The region does not have capacity for the requested model SKU.. This error usually indicates that the region lacks sufficient capacity for the specific SKU requested.

Answer

The API key is invalid.

Answer

The project name is too long.

Answer

The subscription is not registered for the service.

Question 11

Which Azure feature should you use to restrict traffic to your AI Foundry project to specific IP addresses?

Accepted Answer

Azure AI Services 'Networking' (Firewall) settings. Azure AI Services resources (including AI Foundry projects backed by Azure AI Services) expose a Networking (Firewall) configuration blade where you can set 'Selected Networks' and specify allowed IP addresses or CIDR ranges, or disable public access entirely. This is the native, resource-level control for restricting inbound traffic to specific IPs.

Answer

Azure Active Directory Conditional Access

Answer

Network Security Groups (NSG) on the VM

Answer

Azure Front Door WAF

Question 12

An agent orchestration system is timing out because it waits too long for tool output. How can you optimize the infrastructure handling of tool calls?

Accepted Answer

Implement 'Asynchronous' execution patterns for long-running tools.. Implementing asynchronous execution patterns for long-running tools is correct because it allows the agent orchestration to continue processing other tasks or handle timeouts gracefully without blocking on a single tool call. This decouples the tool's execution time from the agent's overall timeout, improving responsiveness and scalability. Increasing CPU or global timeout does not address the fundamental issue of synchronous blocking.

Answer

Increase the 'CPU count' on the hosting server.

Answer

Increase the 'Global Timeout' for the entire agent.

Answer

Reduce the 'System Prompt' size.

Question 13

Which TWO of the following are recommended methods for identifying 'hallucinations' in a RAG system?

Accepted Answer

Using an LLM-as-a-judge to compare against ground truth. Option A is correct because using an LLM-as-a-judge to compare the generated answer against a known ground-truth reference is a standard evaluation technique for detecting hallucinations, as it directly assesses factual consistency and correctness of the output. Option C is correct because comparing the generated output against the retrieval context verifies whether the model's claims are actually grounded in the retrieved documents, which is the core definition of hallucination detection in RAG (faithfulness/groundedness checking). Option B is not appropriate because grammatical errors indicate fluency or language quality issues, not factual fabrication, and a hallucination can be perfectly grammatical. Option D is not appropriate because response speed is a latency/performance metric and has no bearing on factual accuracy. Option E is not appropriate because counting prompt tokens measures input size and cost, not whether the generated content is factually supported.

Answer

Checking for grammatical errors

Answer

Measuring model response speed

Answer

Counting total prompt tokens

Question 14

Which TWO of the following are benefits of using the Azure ML CLI v2 for your MLOps pipelines?

Accepted Answer

Native integration with CI/CD tools. Option A is correct because the Azure ML CLI v2 is designed to be invoked from automation systems such as GitHub Actions, Azure Pipelines, and Jenkins, allowing pipeline steps (e.g., `az ml job create`) to be scripted and triggered as part of CI/CD workflows. Option E is correct because CLI v2 relies on declarative YAML files (such as job, component, and pipeline YAML schemas) to define environments, inputs, outputs, and pipeline steps, which makes configurations versionable and reproducible. Automatic model training (B) is not a benefit of the CLI itself, since training algorithms and jobs must still be defined by the user. No need for Python (C) is false because CLI v2 is a wrapper over the Azure ML Python SDK v2 and many pipeline components are authored in Python. Built-in model visualization (D) is not a CLI v2 feature; visualization is provided through Azure ML studio and related tooling, not the command-line interface.

Answer

Automatic model training

Answer

No need for Python

Answer

Built-in model visualization

Question 15

Which TWO methods are best for debugging an LLM pipeline that fails on complex queries?

Accepted Answer

Inspecting raw input and output tokens via logs.. Option A is correct because inspecting raw input and output tokens via logs reveals exactly what prompt text and context the model received and what it generated, exposing truncation, malformed prompt templates, or unexpected tokenization that cause failures on complex queries. Option B is correct because tracing the request flow through the RAG pipeline shows each stage — query rewriting, embedding, vector retrieval, reranking, and final generation — so you can pinpoint which component drops or corrupts information for complex queries. Option C is wrong because adding API users only increases load and does not diagnose pipeline failures. Option D is wrong because disabling security controls introduces risk without providing any debugging insight. Option E is wrong because deleting and rebuilding the deployment is a destructive, non-diagnostic action that may erase logs and state needed for root-cause analysis.

Answer

Increasing the number of API users.

Answer

Disabling all security controls.

Answer

Deleting the deployment and rebuilding it.

Question 16

Which TWO of the following are effective ways to reduce hallucination?

Accepted Answer

Explicitly instructing the model to say 'I don't know' if the context is insufficient.. Option D is correct because explicitly instructing the model to say 'I don't know' when the context is insufficient gives it a permissible fallback behavior, reducing the tendency to fabricate plausible-sounding but unsupported answers. Option E is correct because supplying ground-truth reference context in the prompt anchors the model's response in verified source material, a technique commonly called retrieval-augmented or grounded generation, which measurably lowers hallucination rates. Option A is incorrect because raising the temperature to 0.7 increases sampling randomness, which typically increases rather than reduces hallucination. Option B is incorrect because removing the system prompt strips away behavioral constraints and role guidance, generally making the model less reliable and more prone to unsupported output. Option C is incorrect because requesting the longest possible response length encourages verbosity and elaboration beyond the available evidence, which tends to increase hallucinated content rather than reduce it.

Answer

Increasing the temperature to 0.7.

Answer

Removing the system prompt.

Answer

Using the longest possible response length.

Question 17

Your team wants to monitor the safety of model outputs. Which THREE steps should you take?

Accepted Answer

Configure 'Diagnostic Logs' to track flagged content.. Option A is correct because configuring Diagnostic Logs in Azure AI services captures flagged content events and prompt/response metadata, giving the team an auditable trail of unsafe model outputs for review and compliance. Option B is correct because enabling Content Safety filters on the deployment applies the harm-category classifiers (violence, hate, sexual, self-harm) to both prompts and completions, actively blocking or annotating unsafe outputs at inference time. Option C is correct because setting up Alerts on metrics such as content violation rates (via Azure Monitor) notifies the team when unsafe output frequency spikes, enabling rapid incident response. Option D does not belong because bypassing filters removes the very safety controls being monitored and increases risk. Option E does not belong because disabling logging eliminates the telemetry needed to detect, investigate, and alert on unsafe outputs.

Answer

Allow all models to bypass filters for performance.

Answer

Disable all logging to save storage space.

Question 18

You are automating model registration using the Azure ML CLI. You need to ensure the registration only happens if the model accuracy is above 0.9. How do you implement this condition?

Accepted Answer

Implement logic in the pipeline to gate the registration step.. You must include logic in your pipeline or script to evaluate the metric (e.g., via a 'PythonScriptStep') and only call the 'az ml model create' command if the condition is met.

Answer

Use the 'condition' parameter in the 'model create' command.

Answer

Use an 'Azure Function' to trigger registration.

Answer

Configure a 'ValidationThreshold' in the registry.

Question 19

Which THREE of the following are common reasons for a model deployment to fail on a Managed Online Endpoint?

Accepted Answer

Missing Python dependencies. Option A (Missing Python dependencies) is correct because a Managed Online Endpoint builds a scoring image from the specified environment; if required packages are absent from the conda/pip specification, the container fails during initialization or at inference time. Option C (Insufficient quota for the VM SKU) is correct because deployment requires available vCPU/memory quota in the target region for the chosen instance type, and exceeding it causes the deployment to fail with a quota error. Option D (Errors in the scoring script) is correct because the scoring script defines init() and run() and is executed by the inference server; syntax errors, bad imports, or incorrect input/output handling cause the deployment or health probe to fail. Option B (Incorrect storage account name) is not a typical cause, since Managed Online Endpoints abstract away storage configuration and do not require the user to specify a storage account. Option E (Slow internet connection) is not a deployment failure cause, as the endpoint runs in Azure's managed infrastructure and does not depend on the client's internet speed.

Answer

Incorrect storage account name

Answer

Slow internet connection

Question 20

What is the purpose of a 'Datastore' in Azure Machine Learning?

Accepted Answer

To manage connectivity to storage services.. In Azure Machine Learning, a Datastore is an abstraction that stores connection information to Azure storage services (Blob, File, Data Lake, SQL, etc.). It lets you reference storage by name in AML without hardcoding credentials or paths. Datastores manage connectivity, not model weights, versioning, or in-memory processing.

Answer

To store model weights.

Answer

To track model versions.

Answer

To process data in memory.