Free Practice Questions for Microsoft AI-103 Exam

Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Apps and Agents Developer/Course Guide/topics:

Use an embedding model . Embedding models convert text into numerical vector representations whose geometric relationships encode semantic similarity.

During indexing, policy documents are normally divided into chunks. Each chunk is submitted to the embedding model and the resulting vector is stored in a vector field in Azure AI Search or another vector store.

At query time, the user ' s natural-language question is submitted to the same embedding model. The resulting query vector can then be compared with the stored document vectors by using a nearest-neighbor similarity algorithm. Semantically similar passages can therefore be retrieved even when the query and policy use different wording.

A large language model or small language model can subsequently use the retrieved text to generate an answer, but the LLM/SLM is not the component specifically intended to create search vectors. An image-generation model creates visual content and is unrelated to textual vector retrieval.

Embedding models are therefore foundational to vector search and many RAG ingestion and retrieval architectures.

Study Guide alignment: configure semantic, hybrid, and vector search for grounding and build retrieval pipelines for internal documents . Microsoft Learn

====================

Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Apps and Agents Developer/Course Guide/topics:

Use the SSML phoneme element. Phonemes provide explicit control over how a word or technical term is pronounced by supplying its phonetic representation.

This is the correct solution when the normal text-to-speech pronunciation rules do not produce the desired pronunciation for specialized terminology, product names, uncommon abbreviations, scientific terms, or organization-specific words.

< say-as > is useful when text should be interpreted according to a known content class—for example, dates, telephone numbers, digits, characters, or other standardized forms. It does not provide the same arbitrary pronunciation control as a phonetic specification.

Prosody controls characteristics such as speaking rate, pitch, and volume but does not redefine the sequence of speech sounds used to pronounce the word.

Training a custom neural voice is substantially more complex and is intended to customize voice identity or related voice characteristics, not simply correct a handful of technical pronunciations. Whisper performs speech-to-text and therefore addresses the opposite speech direction.

The AI-103 Study Guide includes text-to-speech workflows for agentic interactions , making SSML pronunciation controls directly relevant to the Speech portion of the curriculum. Microsoft Learn

====================

The solution does not meet the goal. Protected material detection is intended to identify large language model output that matches known protected text or code, such as copyrighted text, selected web content, song lyrics, articles, recipes, or code. Microsoft describes protected material detection as a control for preventing AI-generated content from reproducing known protected material, not as a control for image safety or prompt injection.

The stated risk has two parts: users can upload unsafe images, and users can embed hidden instructions in images to manipulate the model. Unsafe image uploads require image moderation, because Azure AI Content Safety provides image APIs that detect harmful content across modalities and can support blocking decisions by harm category and severity. Hidden instructions extracted from images are indirect prompt injection or document attacks; Microsoft Prompt Shields are the capability designed to detect user prompt attacks and document attacks, including harmful instructions embedded in third-party content.

Therefore, protected material detection alone does not mitigate either primary risk. Reference topics: Azure AI Content Safety, image moderation, Prompt Shields, document attacks, indirect prompt injection, and protected material detection.

The first statement is No because the code requests only the tags and description visual features. No face-related visual capability is requested. Additionally, facial identification/recognition involves specialized Face capabilities and is not implied merely by calling general Image Analysis.

The second statement is Yes . The code explicitly iterates through image_analysis.tags and prints tag.name together with tag.confidence. The output therefore includes every returned tag and its associated confidence value.

The third statement is also Yes . The statement:

with open(local_image, " rb " ) as image_stream:

opens a path referenced by local_image from the application ' s local file system and provides the resulting binary stream to analyze_image_in_stream. This is different from submitting an externally hosted image URL.

The requested visual features determine which analysis results are generated, while the method used to supply the image determines its source. Here, the input source is a local file and the requested results are image descriptions and semantic tags.

Study Guide alignment: build solutions that analyze visual context and configure applications to produce captions and visual characteristics .

The model must use object detection because the solution needs more than an overall classification. It must identify individual defects, provide their locations, and count individual product samples. Object-detection predictions contain bounding-box coordinates in addition to labels and confidence scores, allowing the application to determine where a defect occurred and whether exactly four product instances are present.

The appropriate domain is General (compact) . The packaging and products are organization-specific and do not map to a more specialized object-detection domain. More importantly, a compact domain supports model export for edge/local inference . This matters because the factory has intermittent internet connectivity; the exported model can execute locally rather than depending continuously on a hosted Azure prediction endpoint.

Microsoft documents compact domains as optimized for deployment outside the Custom Vision hosted service. Exportable formats can include ONNX, TensorFlow, TensorFlow Lite, and other target-specific formats depending on project type and domain.

An ordinary General-domain model is intended for Azure-hosted prediction and cannot satisfy the same offline export requirement. Image classification is also unsuitable because it cannot return defect locations or count independently detected products.

Study Guide alignment: identify objects/components/regions and deploy vision functionality appropriate for edge scenarios . Microsoft Learn

====================

The correct answer is D. prompt shields for documents . The OCR text is extracted from uploaded images and appended as additional context, so it is third-party content rather than a direct trusted user instruction. Microsoft Foundry Prompt Shields distinguish between user prompt attacks , which are malicious instructions directly supplied as user prompts, and document attacks , which are hidden or embedded instructions in third-party content such as documents, emails, webpages, or grounded data. Microsoft also notes that Prompt Shields analyze indirect attacks embedded in input documents or images.

This scenario is an indirect prompt injection pattern: an uploaded screenshot may contain text that says to ignore the system prompt or perform unauthorized actions. Because the OCR output is being passed to the multimodal model as contextual content, the appropriate protection is Prompt Shields for documents, which scans externally sourced context for malicious embedded instructions before generation. Protected material text detects known copyrighted text in model outputs, not prompt injection. Image moderation detects harmful visual content categories, not malicious instructions in extracted context. Prompt Shields for user prompts are less precise here because the risk comes from document-like content derived from the uploaded image. Reference topics: Prompt Shields, document attacks, indirect prompt injection, multimodal safety, OCR-derived context, and Foundry guardrails.

Prompt shields action: Set action to block.

Additional mitigation: Enable Spotlighting.

The correct configuration is to set the Prompt Shields document attack action to block and enable Spotlighting . Prompt Shields in Microsoft Foundry are designed to detect attempts to manipulate model behavior through adversarial input. Microsoft distinguishes document attacks from direct user prompt attacks: document attacks are malicious instructions embedded in third-party content such as documents, webpages, emails, or other externally supplied material. In this scenario, the embedded text inside uploaded screenshots is third-party content and can attempt to override the agent’s instructions. Setting the action to block prevents detected document-attack content from being processed normally, which is required because the goal is prevention rather than passive logging or annotation.

Spotlighting is the additional mitigation because it marks or transforms document content so the model treats it as lower trust than system and user instructions. Microsoft’s Foundry guidance describes Spotlighting as a Prompt Shields subfeature that helps protect against indirect or embedded document attacks by tagging input documents with special formatting to indicate lower trust. A custom blocklist is insufficient for unknown attacks, and OCR alone only extracts the malicious text; it does not mitigate prompt injection. Reference topics: Prompt Shields, document attacks, guardrails, Spotlighting, multimodal safety, and prompt injection defense.

The correct answer is A. an Azure Content Understanding in Foundry Tools analyzer . In Azure Content Understanding, an analyzer is the reusable configuration that defines what content type is processed, which elements are extracted, and how the output is structured. Microsoft’s analyzer guidance states that analyzers define extraction for text, layout, tables, fields, and output formats such as Markdown and JSON. This makes the analyzer the first required configuration step before downstream reasoning or agent orchestration can use the extracted content.

This matches the requirement because the documents contain scanned text, tables, and multicolumn layouts. Content Understanding document analysis is designed to transform unstructured documents into structured, machine-readable output while preserving document structures and relationships. Its Markdown representation converts unstructured documents into GitHub Flavored Markdown while maintaining content and layout for downstream use.

A generative chat completion request or Azure OpenAI Responses API call could reason over extracted content, but it is not the correct first step for OCR, layout preservation, and Markdown conversion. Azure Language focuses on text analysis after content has already been extracted. Reference topics: Content Understanding analyzers, document extraction, OCR, layout analysis, Markdown output, and downstream reasoning.

The correct answer is D. Set the output_content_format=ContentFormat.MARKDOWN value . Azure Document Intelligence Layout API can return extracted document content in Markdown format, preserving semantic structure such as headings, paragraphs, sections, tables, and other layout elements. Microsoft’s Document Intelligence layout guidance shows the Python SDK pattern for analyzing a document with the prebuilt-layout model and setting output_content_format=ContentFormat.MARKDOWN in the begin_analyze_document call. The Markdown output is returned in the top-level content section of the analysis result.

This minimizes development effort because the service produces structure-preserving Markdown directly, rather than requiring custom post-processing to reconstruct sections and table formatting from raw OCR spans. Microsoft’s Markdown output documentation states that specifying Markdown output produces semantically structured content that maintains paragraphs, headings, tables, and other document elements in their proper hierarchy.

Option A only changes validation behavior and does not generate Markdown. Option B requests figures, not structured Markdown. Option C uses an incorrect parameter name; the documented SDK setting is output_content_format, not content. Reference topics: Azure Document Intelligence Layout API, Markdown output, PDF analysis, table extraction, and Foundry Tools document processing.