Free AI-901 practice test

Correct answer: Regression

Forecasting a continuous numeric value such as next month's sales from historical features is a regression problem.

• Anomaly detection: Anomaly detection flags outliers in data; it does not forecast a continuous value.
• Classification: Classification predicts a discrete category, not a continuous sales figure.
• Clustering: Clustering groups similar items without labels; it does not predict a numeric forecast.

Microsoft Learn - Get started with AI on Azure

Correct answer: Optical character recognition (OCR / Read)

OCR - the Read capability of Azure AI Vision - extracts printed and handwritten text from scanned forms and photos.

• Image classification: Image classification assigns a label to a whole image; it does not read the text inside it.
• Object detection: Object detection locates objects with bounding boxes, not the characters of printed or handwritten text.
• Spatial analysis: Spatial analysis tracks people's movement in video, unrelated to extracting text from images.

Azure AI Vision - image analysis overview

Correct answer: Transformer

Modern LLMs are built on the transformer architecture, which uses attention to weigh relationships between tokens regardless of their distance in the sequence.

• Convolutional neural network (CNN): CNNs are typically used for image tasks, not the sequence modelling behind LLMs.
• Recurrent neural network (RNN): RNNs were the pre-transformer approach to sequences; modern LLMs replaced them with transformers.
• Decision tree ensemble: Decision-tree ensembles are classical ML for tabular data, not generative language models.

What is generative AI? (Microsoft Learn)

Correct answer: Reliability and safety

Reliability and safety is the principle that AI systems should perform as intended under expected conditions and respond safely to unanticipated ones; pre-deployment testing across edge cases and adversarial inputs is its canonical practice.

• Fairness: Fairness concerns equitable performance across demographic groups, not robustness to edge cases.
• Privacy and security: Privacy and security concerns protecting data, not how the system behaves under unanticipated inputs.
• Inclusiveness: Inclusiveness concerns designing for diverse users, not adversarial-input testing.

Microsoft Responsible AI principles

Correct answer: Text-to-speech

Reading text aloud in a natural voice is text-to-speech (speech synthesis). Speech-to-text transcribes audio, speaker recognition identifies who is speaking, and keyword extraction is a text capability.

Text to speech (Azure Speech)

Correct answer: An agent with a knowledge tool that retrieves grounded chunks from a connected data source

Foundry agents use tools to ground responses; a knowledge / retrieval tool fetches relevant chunks from your data and the model answers using them, reducing hallucination. The closest distractor is Document Intelligence (option 1) - it extracts structured fields from documents at ingestion time, but it is not a runtime retrieval layer for answering questions, so it cannot replace a knowledge tool here.

Azure AI Foundry - agents overview

Correct answer: A configurable processing unit that defines how content is analysed and what information is extracted

An analyzer is Content Understandings core building block: a reusable configuration defining how input is processed and which fields/outputs are extracted. It is not a billing meter, a VM size, or an identity role.

Content Understanding analyzer reference

Correct answer: The Model catalogue inside the Foundry portal

The Foundry portal includes a Model catalogue (Microsoft, Meta, Mistral, third-party, and your own fine-tunes) where you can compare benchmarks, view safety evaluations, and deploy with one click.

Foundry model catalog overview

Correct answer: from azure.ai.projects import AIProjectClient

AIProjectClient lives in the azure-ai-projects package, so the documented import is 'from azure.ai.projects import AIProjectClient'. This is the unified entry point for calling a Foundry project endpoint.

• from azure.ai.inference import AIProjectClient: azure-ai-inference provides ChatCompletionsClient for a single model endpoint, not AIProjectClient, so this import does not resolve to the project client.
• from azure.ai.ml import AIProjectClient: azure-ai-ml is the Azure Machine Learning v2 SDK and exposes MLClient, not the Foundry AIProjectClient.
• from azure.identity import AIProjectClient: azure-identity only supplies credential classes such as DefaultAzureCredential; AIProjectClient is not defined there.

Get started with Microsoft Foundry SDKs and Endpoints

Correct answer: Retrieval-augmented generation (RAG)

Retrieval-augmented generation (RAG) is the pattern that combines search with the language model: it retrieves relevant content from your own data and includes it in the model input, so responses are grounded in your data rather than only the model's training data. The closest runner-up, fine-tuning, changes model behaviour or style but does not inject fresh source content into each prompt.

• Fine-tuning the base model on the policy documents: Fine-tuning changes the model's behaviour or style and must be retrained to absorb new content; it does not retrieve current documents into each prompt.
• Increasing the temperature parameter on the model: Temperature only adjusts the randomness of the output and supplies no external knowledge to ground the response.
• Prompt caching to speed up repeated requests: Prompt caching is a latency and cost optimisation for repeated prompts; it adds no grounding data from your documents.

Retrieval augmented generation (RAG) and indexes in Microsoft Foundry

Correct answer: Where the model's inferencing data is processed, how you pay, and the performance characteristics

Choosing a deployment type sets three things: where data is processed (global, data zone, or single region), how you pay (pay-per-token vs reserved capacity), and performance characteristics (latency variance, throughput limits). It has nothing to do with portal theming, the model's implementation language, or user access counts.

Understanding deployment types in Microsoft Foundry Models

Correct answer: Both the user prompt (input) and the model completion (output)

The Foundry content filtering system runs both the prompt and the completion through its classification models, so the default filter attached to a deployment screens content in both directions, not just the output.

• Only the model completion (output) - prompts are never inspected: Captures only the output half; the system explicitly evaluates the input prompt too and returns a 400 when a prompt is filtered.
• Only the user prompt (input) - completions are never inspected: Inverts the gap - completions are screened as well, with finish_reason set to content_filter when output is filtered.
• Only stored training data used to fine-tune the model: Confuses an inference-time guardrail with data preparation; content filters act on prompts and completions, not on fine-tuning datasets.

Content filtering for Microsoft Foundry Models