AI-300 sample exam questions

Managed online endpoints can host several deployments, and traffic splitting lets you send a small share to the new version, validate it, then increase. Replacing the endpoint or routing all traffic risks users, and batch endpoints are not for real-time use.

An Azure Machine Learning registry shares models, components, environments and data across workspaces with versioning and central governance, without copying. Git sync, pipeline components or a single environment do not provide cross-workspace asset sharing.

Generate an SSH key pair on the compute instance, display the public key with cat, and add it to GitHub so the terminal can authenticate. Then run git clone in the shared Users file system. The private key always stays on the compute and is never uploaded.

ThresholdOptimizer is a Fairlearn post-processing algorithm that adjusts decision thresholds on an existing model's predictions, so no retraining is needed. ExponentiatedGradient and GridSearch are reduction methods that retrain, and CorrelationRemover is preprocessing.

With multiple deployments behind an endpoint, shifting traffic back to the previous working deployment restores service fastest. Deleting and redeploying is slow, and changing auth type or compute size does not fix errors from a bad release.

A Bandit policy ends runs whose primary metric falls outside an allowed slack from the best run. Setting slack_factor to 0.25 expresses that gap as a ratio. Median stopping compares against running averages, and truncation selection cancels a fixed lowest-performing percentage of runs.

Among the options, gpt-4o is the model that accepts text and image input and supports vision fine-tuning in Foundry. davinci-002 and gpt-35-turbo are text-only, and gpt-4 does not support image fine-tuning.

Prompt variants run in prompt flow and compare in the evaluation view, which records inputs, outputs, token usage and latency for each run, supporting manual comparison. The solution meets the goal.

Bicep templates describe the desired Azure resources declaratively as infrastructure as code. A GitHub Actions or Azure Pipelines step can then run Azure CLI commands such as az deployment group create to apply them. The Python SDK and direct CLI resource commands are imperative.

The Fluency evaluator measures grammatical correctness and readability of the language. Coherence measures logical flow, textual similarity compares to ground truth, and groundedness checks support from source context.

Observability metrics give continuous visibility into live behavior such as latency, errors, and timeouts. An evaluation workflow runs quality and safety evaluators against test data before release, which catches regressions. Diagnostic logs help investigate specific problems, and blocklists only filter terms.

The content safety evaluators include hate and unfairness, which scores whether output is hateful, so they fit validation before exposure. Protected material, groundedness and indirect attacks check other risks.

Choosing a model suited to the task (conversational reasoning), with multimodal support when text and image analysis is required, and with enterprise governance controls for regulated data, matches the requirements. Largest or smallest by default ignore fit and cost-benefit.

A Git repository records who changed each prompt and when, giving traceability across contributors. When applications reference the agent's latest published version, publishing a new prompt takes effect for them without redeploying. Blob storage and datastores offer no contributor history or review workflow.

Tuning chunk size and overlap and adding a re-ranker both improve retrieval without changing the embedding model and at modest cost. Raising token limits increases cost, and changing vector length alters the embedding setup.