CertKata

Microsoft Certified: Azure AI Fundamentals — Free Essentials

Key facts distilled from CertKata's AI-901 question bank. Free to read — no account needed.

Projects, deployments and endpoints

Microsoft Foundry is the platform this exam is built around, and a project is what organises the resources and deployments inside it. A catalog model cannot be called until it is deployed: deployment is what creates a named, callable endpoint, and it carries configuration such as content filters. One deployment can serve several consumers within its capacity, and separate deployments exist when capacity or configuration needs differ - not one per application by default.

The six responsible AI principles

Microsoft names six: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Fairness is about systematically different outcomes for groups of people. Reliability and safety covers consistent, safe behaviour including under unexpected conditions. Privacy and security covers protection of the data used by and produced by the system. Inclusiveness means it works for people of all abilities and backgrounds. Transparency is understanding; accountability is answering.

Hallucination, and grounding as the answer

A hallucination is confidently stated content that is wrong - fabricated citations are the classic case. It happens because a model knows only what was in its training data up to the cutoff, and generates plausible text regardless. Grounding is the mitigation: supply the trusted content in the request and instruct the model to answer from it. Retrieval keeps answers current without retraining, which is how recent events and private data reach a model at all. Unverified output is framed as provisional.

Tokens and the context window

Models read and produce text as tokens, and both directions are usually billed. The context window is the token limit for one interaction, and it bounds input plus output together - a long document, a long conversation history, or a generous output limit all draw on the same budget, and anything beyond it is truncated. That is why history has to be managed rather than accumulated, and why a low maximum-output setting shows up as answers that stop mid-sentence.

Content Understanding

Content Understanding is the extraction service, and it spans documents, images, audio and video - handwriting and varied layouts included, which is the reason to use it rather than parse the files yourself. It pulls defined fields out of forms and information out of recordings. What it does not do is generate. That is the boundary the exam tests: Content Understanding extracts, generative models compose. Extracted data is usually the input to a downstream workflow rather than the end product.

This is a preview. Get the full AI-901 cheat sheet — every domain, bundled with the practice question bank.

Sign up for the full version