AI-901 Reading Companion

Microsoft groups responsible AI into six principles. The exam describes a situation and asks which principle it concerns. Learn the six names and the one-line idea behind each.

How generative AI models work

A large language model (LLM) is trained on a very large body of text. It breaks text into tokens (word pieces) and represents meaning as numeric vectors called embeddings. At its core the model does one thing: it predicts the next token, one token at a time, based on probability. Modern models use the transformer architecture, which uses attention to weigh which earlier tokens matter most.

• Models are stateless. They do not remember past calls. The full conversation history is sent again on every request, and that history counts against the token limit.
• The amount of text a model can consider at once is the context window, measured in tokens (it covers your prompt plus the response).
• Models can hallucinate (produce confident but wrong content). To improve accuracy you ground the model in trusted data using retrieval-augmented generation (RAG), not by lowering temperature.
• Fine-tuning further trains a base model on your own examples to specialise its behaviour. Multimodal models accept more than text — images, audio, or both.

Choosing an appropriate model

Pick the model by what the task needs, then by cost, speed, and context size.

Deployment options and configuration parameters

In Microsoft Foundry you deploy a model from the model catalog, which gives you an endpoint to call. Two broad deployment styles:

• Serverless / standard (pay-as-you-go) — no infrastructure to manage, you pay per token. Fast to start, good for variable or low traffic.
• Managed compute (dedicated) — you reserve compute for predictable, high-volume workloads.

You may also choose a region or data-zone for data residency. After deployment you tune inference parameters:

Naming the workload (1.3.1)

The exam gives a scenario and asks which workload it is. Decide by what the system does with the input, not by what the input is.

Text analysis techniques (1.3.2)

Other capabilities in Azure AI Language: language detection (which language), PII detection (find and redact personal data), entity linking, conversational language understanding (CLU) for intent and entities, and custom question answering for a Q&A knowledge base.

Speech (1.3.3)

• Speech-to-text (speech recognition) — audio in, text out. Real-time or batch.
• Text-to-speech (speech synthesis) — text in, spoken audio out, using neural voices.
• SSML (Speech Synthesis Markup Language) — XML markup to control pitch, pace, pauses, and switch voices.
• Speech translation — spoken word in one language to another.

Computer vision and image generation (1.3.4)

Image generation is the reverse: a generative model (for example DALL-E) creates a new image from a text prompt. Recognising versus creating an image is a common exam split.

Extracting information from text, images, audio, and video (1.3.5)

When the goal is structured fields out of mixed content, the workload is information extraction. Across documents, images, audio, and video this is the job of Azure Content Understanding. Forms and documents specifically can also use Azure AI Document Intelligence.

The Foundry mental model

Microsoft Foundry is the unified portal at ai.azure.com for building AI solutions. You work inside a Foundry project, deploy models from the model catalog, experiment in the playground, and call deployed models from code through the Foundry SDK.

System prompt vs user prompt (2.1.1)

Deploy and interact in the portal (2.1.2)

From the model catalog you select a model, deploy it, and test it in the chat playground before writing any code. Deploying gives you an endpoint and a deployment name.

Lightweight chat client with the Foundry SDK (2.1.3)

The build lessons call Foundry from Python using the azure-ai-projects package and authenticate with DefaultAzureCredential from azure.identity — keyless sign-in through Microsoft Entra after az login. The only configuration value you supply is the project endpoint.

Single agent in the portal (2.1.4) and its client app (2.1.5)

An agent in Foundry is exactly three parts:

The tools you can attach are a frequent exam topic:

A client app for an agent uses the azure-ai-agents capabilities of the Foundry SDK, again with DefaultAzureCredential.

Text analysis app (2.2.1)

You build a lightweight app that analyses text — for example classifying sentiment, pulling key phrases, or extracting entities — using a deployed model in Foundry or Azure AI Language capabilities. Match the requested task to the right capability (sentiment for tone, key phrases for topics, NER for names).

Respond to spoken prompts with a multimodal model (2.2.2)

A multimodal model (for example GPT-4o) can take audio directly as input and respond, without a separate transcription step. This preserves tone and intent that a text-only pipeline would lose.

Azure Speech in Foundry Tools (2.2.3)

For dedicated, enterprise speech — call-centre transcription, custom neural voices, fine SSML control — you use Azure AI Speech through Microsoft Foundry Tools: speech-to-text for transcription and text-to-speech with SSML for synthesis.

Interpret visual input with a multimodal model (2.3.1)

Send an image to a multimodal model and ask about it — describe the scene, answer a question about the picture, read a chart. Same idea as audio: the model takes the image as input alongside your text prompt.

Create visual outputs with generative models (2.3.2)

Use an image-generation model (for example DALL-E) to produce a new image from a text prompt — marketing visuals, mockups, illustrations.

Build a vision app (2.3.3)

Combine the two: an app that interprets images with a multimodal model, generates images when needed, or calls Azure AI Vision for classic capabilities such as OCR, image classification, and object detection.

Azure Content Understanding, available through Microsoft Foundry Tools, extracts structured information across four content types. You define an analyzer (the fields or schema you want), run content through it, and get structured results back.

Objective 2.4.4 ties it together: build a lightweight app whose job is information extraction using Content Understanding.

These are Microsoft's own free hands-on exercises — the authoritative labs for this exam. Do them in order: lab 0 creates the Foundry project that every later lab reuses. Tick each one as you finish (saves in this browser).

A credit or debit card is required for identity verification on the free account, but you are not charged unless you manually upgrade to pay-as-you-go.