Microsoft AI-300 Online Practice Questions and Exam Preparation

Microsoft AI-300 Online Questions & Answers

• Question 41:

  HOTSPOT

  You manage a Retrieval-Augmented Generation (RAG) system that uses Azure AI Search to retrieve documents from an indexed knowledge base.

  The system must support the following retrieval requirements:

  1. Queries that include exact policy identifiers must return matching documents even when semantic similarity is low.

  2. Natural-language questions must prioritize semantically relevant documents even when keywords are not an exact match.

  You need to configure the retrieval approach to meet the requirements.

  How should you configure the retrieval behavior for each requirement? To answer, select the appropriate options in the answer area.

  NOTE: Each correct selection is worth one point.

  

  

  ---

  The image maps exact identifier matching to keyword-style retrieval and natural-language relevance to semantic or vector retrieval. Hybrid search combines these approaches so identifiers can be matched precisely while conceptual queries still retrieve relevant passages. Other single-mode retrieval choices would miss one side of the requirement and reduce RAG quality.

• Question 42:

  A team manages an Azure Machine Learning workspace and deploys a model to an endpoint. A deployed online endpoint shows inconsistent response times during periods of high traffic.

  You need to identify potential performance degradation.

  Which three metrics should you monitor? (Choose three.)

  A. Feature count
  B. Requests per minute
  C. Connections active
  D. Dataset size
  E. Request latency
  B. Requests per minute
  C. Connections active
  E. Request latency

  ---

  Explanation

  Inconsistent online endpoint response times are handled by monitoring endpoint performance, reviewing deployment behavior, and adjusting capacity or autoscaling settings. Those actions address latency, throughput, and deployment health. Other alternatives that only change training data or model registration would not directly improve live endpoint responsiveness under high traffic.

• Question 43:

  You are fine-tuning a base language model to analyze customer feedback. You label examples of support tickets.

  You must improve classification accuracy by configuring and fine-tuning the base model in Microsoft Foundry.

  You need to configure and run fine-tuning.

  What should you do first?

  A. Use prompt flow to generate multiple prompt templates for evaluation.
  B. Deploy the base model to an online endpoint before starting fine-tuning.
  C. Enable tracing for all inference calls in the evaluation pipeline.
  D. Format the dataset as a JSONL file with prompt-completion pairs and upload the file.
  C. Enable tracing for all inference calls in the evaluation pipeline.

  ---

  Explanation

  Classification accuracy for labeled support-ticket examples is improved by fine-tuning with task-specific labeled examples that teach the model the target categories. Retrieval or monitoring alone does not change the model behavior for the classification task. Other alternatives may help evaluation or operations, but the requirement asks for model adaptation through fine-tuning.

• Question 44:

  DRAG DROP

  A team iterates prompts used by a generative AI agent. The team must support internal review before releasing changes.

  The team must:

  1. Track prompt changes with a clear history for audit and rollback.

  2. Compare prompt variants in parallel without affecting the prompt used in the production environment.

  You need to select the appropriate source control approach for each requirement.

  What should you use for each requirement? To answer, move the appropriate source controls to the correct requirements. You may use each source control once, more than once, or not at all. You may need to move the split bar between panes or scroll to view content.

  NOTE: Each correct selection is worth one point.

  Select and Place:

  

  

  ---

  The source-control image maps prompt change tracking to Git and prompt evaluation to a review or comparison workflow before release. Git records history and ownership, while evaluation checks behavior before changes reach users. Other choices may store prompts, but they do not satisfy both traceability and internal review requirements

• Question 45:

  A Retrieval-Augmented Generation (RAG) solution returns incomplete answers because relevant content is inconsistently retrieved from the knowledge source.

  You need to improve RAG accuracy without changing the embedding model currently in use. You need to achieve this goal while minimizing operational costs.

  Which two actions should you perform? ( Choose two.)

  A. Tune chunk size and overlap to match content structure.
  B. Implement an optimized re-ranker.
  C. Increase token limits for all requests.
  D. Optimize the length of embedding vectors.
  A. Tune chunk size and overlap to match content structure.
  B. Implement an optimized re-ranker.

  ---

  Explanation

  Incomplete RAG answers often come from retrieval configuration, so tuning similarity thresholds and chunking strategy directly addresses inconsistent retrieval of relevant content. These settings affect recall and context quality before generation occurs. Other actions such as changing only the final prompt or deployment capacity would not fix missing source passages in the retrieved context.

• Question 46:

  A team is deploying machine learning models to a production inference endpoint in Azure Machine Learning.

  The team requires a safe way to validate a new model version without disrupting existing users.

  You need to recommend a deployment strategy for controlled testing of a new model version.

  What should you configure?

  A. traffic splitting between deployments
  B. the model asset version in the registry
  C. deployment to a separate staging endpoint
  D. an evaluation script in Azure Machine Learning
  A. traffic splitting between deployments

  ---

  Explanation

  Traffic splitting is the safe deployment strategy because it lets the new model receive a controlled share of production traffic while the existing deployment remains available. That supports validation without disruption. Model registry settings, staging-only endpoints, or evaluation scripts can support the workflow, but they do not control live production traffic as required.

• Question 47:

  HOTSPOT

  You train a model in Azure Machine Learning.

  You plan to capture experiment details for later comparison.

  The training code must log parameters and metrics for each run.

  You review the following training script.

  

  You need to verify whether the training script meets the experiment tracking requirement.

  For each of the following statements, select Yes if the statement is true. Otherwise, select No.

  NOTE: Each correct selection is worth one point.

  

  

  ---

  The code in the image starts an MLflow run and logs parameters, metrics, and the trained model, so it supports later experiment comparison. The experiment name groups runs, but it is not the artifact path for the logged model. The takeaway rule is that MLflow run logging records experiment evidence, while artifact location and registration are separate lifecycle concerns.

• Question 48:

  A team develops and manages a conversational assistant by using Microsoft Foundry.

  The team must be able to validate that the assistant does not produce hateful responses before the application is exposed to any users.

  You need to evaluate the model output for hateful responses as part of a repeatable validation process.

  Which evaluator should you configure first?

  A. Protected material
  B. Groundedness
  C. Indirect attacks
  D. Content safety
  D. Content safety

  ---

  Explanation

  Risk and safety evaluation is the correct mechanism for validating that a conversational assistant does not produce harmful or hateful responses. Groundedness or latency metrics measure different quality dimensions. The requirement is safety validation before release, so harmful-content detection and safety evaluation are the appropriate controls.

• Question 49:

  You have a Microsoft Foundry project.

  You plan to use the Microsoft Foundry portal to fine-tune a base Azure OpenAI Service model that can accept both text and images as input.

  You need to choose the suitable model.

  Which model should you choose?

  A. davinci-002
  B. gpt-4o
  C. gpt-35-turbo
  D. gpt-4
  B. gpt-4o

  ---

  Explanation

  Fine-tuning a multimodal base model requires a model and fine-tuning path that supports the input modalities used by the examples, including images. A text-only model or unsupported tuning path would not meet the requirement. The other alternatives may support prompt testing or deployment, but they do not enable the requested multimodal fine-tuning workflow.

• Question 50:

  DRAG DROP

  A real-time endpoint is deployed in Azure Machine Learning to serve predictions to a web application.

  Users report intermittent failures and unexpected responses when calling the endpoint.

  You need to identify the appropriate troubleshooting action for each reported issue.

  Which troubleshooting action should you perform for each issue? To answer, move the appropriate troubleshooting actions to the correct issues. You may use each troubleshooting action once, more than once, or not at all. You may need to move the split bar between panes or scroll to view content.

  NOTE: Each correct selection is worth one point.

  Select and Place:

  

  

  ---

  The answer maps failures after deployment updates to reviewing deployment logs and incorrect responses to validating the scoring script and model version. Logs expose runtime errors and deployment startup failures, while incorrect prediction behavior often comes from model or scoring-code mismatch. Other actions, such as restarting or autoscaling, can help availability but do not directly diagnose those specific requirements.