AI-901 — frequently asked questions

SpeechConfig handles the service connection (key/region), while AudioConfig specifies the audio input/output device or stream.

No, SpeechSynthesizer requires a SpeechConfig instance to initialize and connect to the Azure Speech service.

System instructions set the persistent role and rules for the agent, while user prompts provide the specific input or task for a single interaction.

No, temperature only affects randomness and creativity in word choice. Personality and role are defined strictly through system instructions.

No. Authentication and authorization are handled by external identity providers like Microsoft Entra ID or via API keys, not by the text within a system prompt.

No. TPM is a service quota defined by the cloud provider to manage load. While prompts can influence response length, they do not configure or reduce the TPM limit itself.

OCR only converts images to text. It cannot inherently understand or map that text into a structured schema with nested fields required for expense calculation.

Azure AI Search indexes and searches data. It does not extract or structure the initial document content; it relies on other services to provide the structured data first.

Because invoice analysis is an asynchronous background task. You must use a poller to wait for the server-side processing to finish before the data is available.

It synchronously waits for the long-running operation to complete and then returns the final result object containing the extracted information.

The model name (e.g., gpt-4.1-mini) is the base architecture. The deployment name (e.g., my-mini-gpt) is the unique alias you assign to that model instance in your specific Azure resource.

The endpoint URL is found in the Azure Portal under the Azure OpenAI resource properties. It typically follows the format https://<resource-name>.openai.azure.com/

Models are pre-trained on massive datasets. Retraining them for every prompt would be computationally impossible and slow down responses.

A model generates content like text or images. An agent uses reasoning to plan steps, make decisions, and execute actions using tools.

Document Intelligence extracts text and structure from documents, while Computer Vision analyzes visual elements like objects and scenes in images.

No, Computer Vision lacks the advanced layout recognition and table extraction capabilities required for complex document processing tasks.

Transparency focuses on openness and visibility, such as naming teams and explaining decisions. Accountability focuses on governance, taking responsibility for outcomes, and having mechanisms for audit and redress.

No. Naming the team primarily ensures Transparency by identifying who is involved. Accountability requires a broader framework including governance, documentation, and monitoring to ensure the team can be held responsible.

Accountability assigns ownership of outcomes to people, while Transparency only requires disclosing how a system works and its limits. The sentence stresses reviewers taking responsibility, which is Accountability.

Expect Fairness, Reliability and safety, Privacy and security, Inclusiveness, and Transparency alongside Accountability. Only Accountability describes people staying answerable for AI-driven decisions.

Classification only assigns labels or categories to text that already exists, such as sentiment or intent. It cannot author new sentences, so it cannot turn a product description into an email draft.

You would add a speech recognition workload to transcribe the audio first, but the drafting step itself still requires generative AI to write the email text.

Fairness focuses on eliminating bias and ensuring equitable outcomes for different groups. Inclusiveness focuses on making AI accessible and usable for everyone, including those with disabilities.

Microsoft defines Fairness as treating all people fairly and avoiding affecting similarly situated people differently, which includes mitigating bias in data and models.

Inclusiveness focuses on accessibility for all users, including those with disabilities. Fairness focuses on preventing bias and ensuring equitable outcomes across demographic groups.

Encrypting data falls under the Privacy and Security principle, which protects user information. Inclusiveness is about making the product usable by everyone regardless of their abilities or language.

client.images.generate creates visual media from text prompts, while client.chat.completions generates text-based conversational responses.

For image generation tasks, use models like dall-e-3 or gpt-image-1 depending on availability, ensuring the method matches the image API.

No, MCP is not mandatory. It is an optional protocol that allows agents to connect to external tools and data sources, but agents can function using other methods.

Yes, you can use MCP if you need standardized tool integration, but it is not a prerequisite for creating or running basic agents in Azure AI Foundry.

Computer Vision analyzes and interprets existing images (e.g., detecting objects). Multimodal Generative AI creates new content (e.g., generating an image from a text description).

Yes, some advanced multimodal models can understand images, but their primary distinction in this context is the ability to generate new content rather than just classify existing data.

Speech workloads convert audio to text (transcription), while NLU workloads analyze existing text to understand intent, entities, and sentiment.

No. Key phrase extraction analyzes text documents to find significant terms. You must first transcribe the audio using a Speech workload before applying text-based analytics.

Key phrase extraction is a Natural Language Processing (NLP) function handled by Azure AI Language, not Azure AI Vision.

While it can detect objects, identifying specific breeds usually requires custom models or specialized services, not the standard Foundry thumbnail tool.

NER identifies specific proper nouns like names and locations, while Key Phrase Extraction pulls out general topic words that summarize the text.

No, Azure AI Language provides pre-trained models for NER that work out-of-the-box for standard entity types like persons and organizations.

Custom NER requires training and maintenance. PII detection is a pre-built, optimized service specifically for common sensitive data types like names and phones.

Yes, Azure AI Language PII detection includes built-in models that recognize and can redact phone numbers along with other personal identifiers.

It accepts the image either as inline base64-encoded image data embedded in the JSON payload or as a publicly accessible image URL. No other transport, such as CSV or MP4, is used for this image input.

An SAS token is only a time-limited credential that grants access to stored blobs; it does not contain image content. It might accompany a URL, but the image itself is passed as base64 data or a URL.

Overall accuracy is an aggregate across the whole dataset, so poor performance on a small underrepresented group is diluted. Fairness must be measured per subgroup, not on the total score.

Slice evaluation data by sensitive attributes such as demographic groups, then compare error rates and outcome metrics across those slices to reveal disparities, per Microsoft's fairness measure guidance.

A separate request forces an extra network round trip before the model can see the image, so total latency increases instead of decreasing. Combining both items in one user message lets the model process them simultaneously.

The system message sets standing instructions and role behavior, while the user message carries the actual prompt content. Image references belong with the user's request so they are interpreted as part of that turn.

Document analysis is a long-running operation, so the SDK returns a poller that must be awaited; results are only available after processing finishes.

Headers only carry operation metadata such as the operation-location URL; the actual extracted fields come from polling the operation until it completes.

Image generation creates new images from prompts, while image analysis interprets existing images to extract data or labels.

No, OCR only extracts text from images. It cannot create visual content.

Session history logs past interactions for review but does not contain the live API endpoint or security tokens needed to send new requests.

No, the display name is for human identification in the UI; the API uses a unique resource ID or specific endpoint URL for routing.

Yes, but via the 'Speech Translation' feature, not 'Text-to-Speech'. TTS is strictly for converting existing text to audio in the same language.

TTS converts text to speech in one language. Speech Translation converts speech from one language to speech in another, or text to translated speech.