AI-901 Cheat Sheet: Models, Agents, and Foundry Cheat Sheet
Cheat sheet: compare prompts, clients, agents, text and speech tasks, visual workflows, and structured extraction for AI-901 practice.
Use this reference when reviewing an IT Mastery explanation. It summarizes the distinctions behind common decisions; the blueprint page shows the complete preparation map. Use IT Mastery for varied AI-901 practice, then use these recall checks to identify what needs another session. Answer aloud or write a sentence before opening each check. Mark an answer for review if you guessed or could not explain the deciding fact. These are original IT Mastery review prompts, not official Microsoft questions, copied live-exam content, or exam dumps. Use the reference tables on this page when you need a definition or comparison.
Use the tables for a quick pre-exam check. Expand a topic’s notes for explanations, examples, and additional distinctions.
Models and configuration
| Concept | Practical distinction |
|---|---|
| Token | A unit a model processes; it need not equal a whole word. |
| Embedding | A numerical representation useful for comparing meaning; similarity is not proof of identical facts. |
| Attention | A mechanism that relates parts of the input in context. |
| Model capability | What inputs and tasks a model supports; assess it against the actual requirement. |
| Deployment | A configured model target. Its operating constraints differ from a single request’s settings. |
| Temperature | Where supported, affects output variation; a low value does not guarantee correctness or exact repetition. |
| Maximum output tokens | Limits generated output length; it does not add missing source facts. |
| Grounding | Supplies relevant source context. Retrieval-augmented generation uses retrieval and prompting, not retraining. |
Prompts, format, and evidence
| Prompt element | What it supplies |
|---|---|
| System prompt | Reusable application guidance: role, behavior, boundaries, and expected response format. |
| User prompt | The current task and its input, such as the ticket to summarize. |
| Output requirements | The fields, structure, and missing-information behavior the application needs. |
Specify the required output, provide relevant source facts, and validate the result. Valid JSON establishes that the text parses; it does not establish that its values are accurate or supported. Instructions also need testing. Microsoft’s system message guidance covers roles, formats, fallback behavior, and their limits.
Application and agent checks
| Part | What to check |
|---|---|
| Endpoint | The interface expected by the client; a Foundry project endpoint is distinct from an Azure OpenAI endpoint. |
| Credential | The authentication method supported by that operation, with permission to access the intended resource. |
| Request | The actual input, instructions, target reference, and supported settings sent by the application. |
| Agent definition | Reusable model, instructions, and available tools; a direct model call does not inherit them. |
| Conversation context | Relevant history or a supported reference carried between turns; it is separate from the agent definition. |
| Response handling | Read the required text, audio, image, or fields from the actual returned structure. |
| Asynchronous operation | Submit, check completion, then retrieve or consume the result as the API requires. |
A REST body depends on the operation: it may contain JSON, audio bytes, or another required format. An SDK builds the underlying requests, but the application still needs correct input and result handling.
Text and audio tasks
| Required result | Capability or distinction |
|---|---|
| Main topics in text | Key-phrase extraction. |
| Typed names, places, or organizations | Entity recognition. |
| Overall opinion | Sentiment analysis. |
| Opinion about a particular target | Opinion mining adds target/assessment detail while retaining overall sentiment. |
| Text with sensitive spans removed | PII redaction; inspect the returned redacted text and validate the required coverage. |
| Shorter account of source material | Summarization; check retained meaning and required details. |
| Written words from speech | Speech recognition; transcription alone does not fulfill the spoken request. |
| Audio from written words | Speech synthesis; select the required voice and output destination. |
| Immediate results from streamed speech | Real-time recognition; the stream may come from a microphone or a file. |
| Deferred transcription of stored recordings | Batch transcription when immediate results are unnecessary. |
| Managed spoken interaction | Voice Live; the selected model or agent determines whether it uses native audio or separate recognition and synthesis stages. |
Notes and examples
Use Microsoft’s Voice Live documentation when checking a particular model configuration.
Visual tasks and information extraction
| Task | Key distinction |
|---|---|
| Image classification | Assigns an overall image label. |
| Object detection | Identifies and locates objects. |
| Semantic segmentation | Assigns classes at pixel level. |
| Multimodal image interpretation | Combines supported image input with a text instruction. |
| Image creation or editing | Check whether the operation creates a new composition or targets changes to a source; use of an existing image alone does not distinguish the two. |
| Video generation | Produces visual content over time; follow the job’s completion and retrieval workflow. |
| OCR | Recognizes printed or handwritten text; results can include word and line locations. |
| Field extraction | Maps source evidence to named business values and relationships. |
| Content Understanding analyzer | Defines processing and the required field schema; compare prebuilt coverage with the actual output needs. |
Notes and examples
Recognizing words and their positions does not by itself establish invoice-line relationships or named business fields. Microsoft’s OCR overview distinguishes text recognition from further document analysis. The extraction walkthrough shows how an application can lose those relationships even when it has read the individual values.
Before integrating extracted results, check mandatory fields, types, nested relationships, collections, and source evidence. A successful request or well-formed JSON does not prove that the result is complete or correct. Current Content Understanding confidence and source-grounding features are associated with document analyzers; do not assume equivalent confidence fields for every modality. See Microsoft’s document field extraction guidance .
Responsible AI checks
Evaluate representative groups for fairness, test failure behavior for reliability and safety, minimize and protect personal data, provide accessible alternatives, disclose relevant limitations, and assign responsibility for outcomes. Choose the safeguard that addresses the stated risk rather than applying one generic control to every case.
Continue with applied practice
Recall AI concepts and capabilities
Model selection. A model leads a general text benchmark, but the application must interpret images. What evidence do you need before choosing it?
Check model selection
Verify that the model and deployment support the required image input and task, then check the application's other constraints. A text benchmark alone does not establish visual capability. If both candidates support the task, compare relevant quality, latency, and operating constraints.
Responsible AI. A service performs well overall but misses substantially more eligible cases for one user group. Would publishing an AI-use notice resolve that finding?
Check the responsible AI decision
No. The finding calls for investigating performance across representative groups and addressing the source of the disparity. A notice can support transparency, but it does not correct unequal performance. If the evidence instead showed an inaccessible interface, the needed response would concern inclusiveness.
Generation settings. A response ends before its required conclusion. Why might an output-token limit matter, and why would lowering temperature not establish that the completed answer is factual?
Check generation settings
Inspect whether the request reached its output limit and whether the requested response fits that limit. Where supported, temperature affects variation; it does not supply evidence or guarantee factual accuracy. Increasing an output limit also does not repair missing source context.
Recall lightweight implementation decisions
Prompt construction. Every ticket summary must use the same fields and mark unknown details as missing. Each request supplies a different ticket. Where should the reusable guidance and current ticket go, and does valid JSON prove that the summary is accurate?
Check prompt construction
Put reusable behavior, format, and missing-information guidance in the system prompt or saved agent instructions. Supply the current task and ticket as request input. Then compare the returned values with the ticket: valid JSON can still contain invented details. Test both ordinary and incomplete tickets.
Client data flow. A playground gives the expected answer, but a small Python application displays an empty result. What would you compare before changing models?
Check the client path
Compare the actual input, target deployment, endpoint, authentication, supported settings, and returned structure. Inspect any error first. A successful call may still be displayed incorrectly if the application reads the wrong result field or consumes an asynchronous result before completion.
Agent context. An application invokes the correct saved agent, but a follow-up such as “What about tomorrow?” loses the earlier subject. What information may be missing?
Check agent context
Check the conversation history or supported conversation reference supplied with the follow-up. An agent definition provides reusable behavior and tools; it does not by itself identify which earlier exchange this request continues. Passing unrelated history would also fail the requirement.
Text, speech, and visual input. A client can transcribe a spoken question. What is still needed to answer it aloud? Separately, does writing an image’s filename in a prompt supply its visual content?
Check input and output handling
The spoken interaction still needs an answer based on the question and relevant context, then an audio-output path. A deployed multimodal model or a supported voice workflow may handle these stages differently. For visual interpretation, the request needs supported image input; mentioning a filename in prose does not supply the image.
Information extraction. An application needs action items and owners from a meeting recording. Is a transcript alone a completed result?
Check extraction against the requirement
No. The application must extract the required information into the expected fields and preserve each action item's relationship to its owner. Check the analyzer's support for the source modality, its schema, the returned values, and missing information. Do not invent an owner when the recording does not identify one.
Turn an uncertain answer into the next session
| What stopped you? | Next action |
|---|---|
| An unfamiliar concept | Review its definition, then distinguish it from the nearest alternative without notes. |
| A missed scenario condition | Write the deciding fact and explain how changing it could change the answer. |
| Difficulty following code or results | Trace input → request → operation → result → application use in a short Microsoft example. |
| Remembering an answer but not its reason | Try fresh questions in that area and justify the choices before reading explanations. |
Revisit the uncertain area in a later mixed session. Use the scenario guide for a worked decision and the study plan to organize repeated gaps. Completing these checks is a review aid, not a prediction of your exam result.