Free AI-901 Exam Flashcards

Microsoft Foundry is the workspace for building, testing, deploying, and managing AI apps and agents on Azure. For AI-901, connect Foundry to model deployment, prompt testing, agents, tools, and app integration.

A Foundry project groups the resources for one AI solution, such as deployments, data connections, tools, and evaluation assets. It keeps solution work organized instead of scattering pieces across separate portals.

The model catalog is where you compare available models by capability, modality, cost, and deployment option. Pick the model that matches the task, not just the most powerful model.

Deployment makes a selected model callable by an application. Without a deployment, a model may appear in the catalog but there is no configured endpoint for inference requests.

The chat playground is an interactive way to test a deployed model, prompts, and parameters before writing code. Use it to learn behavior quickly and reduce trial-and-error in an app.

The Foundry SDK lets code connect to a project, call model deployments, and build lightweight clients or agents. It is the programmatic path after a portal prototype works.

Azure AI services provide ready-made APIs for common workloads such as vision, language, and speech. Use them when you need proven capabilities without training a custom model from scratch.

Fairness means an AI solution should avoid systematically disadvantaging people or groups. The consequence is design and testing that look for biased outcomes, not just high overall accuracy.

Reliability and safety focus on consistent behavior and harm reduction under expected and unexpected conditions. Testing, monitoring, fallback behavior, and content filters support this principle.

Privacy and security require protecting training data, prompts, outputs, and connected systems. A useful AI app can still fail this principle if it leaks personal or confidential information.

Inclusiveness means the AI solution should work for people with different abilities, backgrounds, and access needs. Accessibility is not a cosmetic add-on; it affects who can benefit from the system.

Transparency means users and stakeholders can understand what the system can do, what it cannot do, and when AI is being used. It reduces misplaced trust in model output.

Accountability means people and organizations remain responsible for AI behavior and outcomes. A model cannot own a policy decision, approval workflow, or failure response.

Content safety detects and blocks harmful text or images, such as violent, hateful, or self-harm content. It is especially important when users can submit open-ended prompts.

Human-in-the-loop design sends sensitive or uncertain AI outcomes to a person for review. It is useful when a wrong answer could harm a person, business, or legal process.

A generative language model creates text by using learned patterns to predict likely next tokens. It can produce new responses, but fluency does not guarantee factual correctness.

A token is a chunk of text processed by a language model. Prompts, retrieved context, and generated answers all consume tokens, which affects cost, latency, and context limits.

The context window is the amount of text and other input a model can consider at once. If important information falls outside it, the model cannot use that information directly.

Temperature controls response variation. Lower values make output more predictable; higher values make it more diverse but can reduce consistency for factual or formatted tasks.

Max tokens limits how much text the model can generate. It is a guardrail for answer length and cost, but setting it too low can cut off useful responses.

Embeddings turn text or other content into numeric vectors that represent meaning. They make semantic search possible because similar ideas can be found even when wording differs.

Semantic search retrieves content by meaning instead of only exact keyword matches. It is useful when users ask questions in different words than the source documents use.

RAG retrieves relevant source content and adds it to the prompt so the model can answer with grounding. It improves domain accuracy without changing the base model weights.

Use RAG to ground answers in changing or private content. Use fine-tuning when you need the model to learn a repeated style, task pattern, or behavior from examples.

A hallucination is a confident answer that is unsupported or false. Grounding, citations, content limits, and human review reduce risk but do not remove it entirely.

A multimodal model can work with more than one input or output type, such as text plus images. Use it for tasks like answering questions about a screenshot or chart.

Generative AI creates new content, such as text or images. Predictive AI estimates a value, class, or future outcome from patterns in data.

An AI agent uses a model plus instructions and tools to work toward a goal across steps. It differs from a simple chat response because it can decide when to call actions.

Tools let an agent do work beyond text generation, such as searching data, calling an API, or running a workflow. Tool design controls what the agent can actually affect.

Instructions define an agent's role, boundaries, and expected behavior. Clear instructions reduce unwanted actions and make tool use more predictable.

A system message sets high-priority guidance for a chat model, such as role, tone, limits, and safety rules. It is more durable than one-off user wording.

Few-shot prompting includes sample inputs and outputs so the model can follow the pattern. It changes prompt context, not the trained model weights.

Prompt constraints tell the model what format, sources, tone, or boundaries to follow. They make output easier to validate and safer to use in applications.

Sentiment analysis estimates whether text expresses positive, negative, neutral, or mixed feeling. It is useful for reviews and feedback, not for proving factual claims.

Key phrase extraction identifies important terms that summarize what a text is about. It helps tag or scan documents without reading every sentence.

Named entity recognition finds people, places, organizations, dates, and similar entities in text. It extracts structured signals from otherwise unstructured language.

Summarization shortens text while preserving the main meaning. Extractive summaries reuse key source sentences; abstractive summaries restate the ideas in new wording.

Translation converts content from one language to another. It is an NLP workload when translating text and a speech workload when translating spoken audio.

Speech to text transcribes spoken audio into written words. It supports captions, meeting notes, voice commands, and searchable audio archives.

Text to speech turns written text into spoken audio. It is used for narration, accessibility, call center responses, and voice experiences.

Speaker recognition identifies or verifies a person by voice characteristics. It is different from speech recognition, which determines the words being spoken.

Image classification labels the whole image. Object detection identifies individual objects and where they are located, often with bounding boxes.

OCR extracts printed or handwritten text from images and documents. It is the right workload when the goal is to read visible text, not classify the image.

Image generation creates a new image from a prompt. Image analysis interprets an existing image, such as labeling scenes, reading text, or finding objects.

Azure AI Vision supports image analysis tasks such as OCR, image captions, object detection, and visual tagging. Map it to scenarios that start with image or video content.

Azure AI Language supports text analytics such as sentiment, entities, key phrases, language detection, and summarization. Map it to scenarios built around written language.

Azure AI Speech supports speech to text, text to speech, speech translation, and speaker recognition. Map it to scenarios where audio is the main input or output.

Azure Content Understanding extracts structured information from documents, images, audio, and video. Use it when the task is to pull fields, topics, or key details from content.

Document field extraction finds values such as dates, totals, names, and identifiers in forms or invoices. The goal is structured output, not just a text transcript.

Classification predicts a category, such as approve or deny. Regression predicts a numeric value, such as price or demand.