AI-103 sample exam questions

Content Understanding standard mode is the cost-effective option for high-volume processing of single, standalone files such as invoices. Pro mode adds multi-step reasoning, multi-file input and reference data, which is needed to validate documents against each other in Pipeline2.

Not(IsBlank(Local.Var01)) is true only when the variable holds a value; IsBlank alone tests the opposite, and IsEmpty is for tables. In a Send message node, Power Fx inside braces is evaluated, and the variable must be referenced with its Local scope prefix, so {Upper(Local.Var01)} returns the answer in uppercase.

To keep harmful content out of agent runs, the guardrail must cover every intervention point (user input, tool call, tool response and output) with the Block action; Annotate only labels content. Content Safety only needs to read the blobs, so a managed identity with Storage Blob Data Reader is least privilege, unlike account keys or a Contributor role.

When a custom speech model expires, Azure Speech does not fail or delete it; recognition requests to the custom endpoint automatically fall back to the most recent base model for the same locale. You should retrain on a newer base model before expiry to keep the custom accuracy.

A large language model (such as GPT-4.1 or a reasoning model) is needed for deep multi-step reasoning over long, retrieved contexts and detailed natural language generation in a RAG agent. Small language models trade reasoning depth for cost, multimodal capability is not required, and key phrase extraction cannot generate answers.

Instrument the Python service with OpenTelemetry (the Azure AI/Foundry SDK tracing instrumentation) and export the spans to Application Insights, which is the trace store Foundry uses for agent tracing and latency and exception views. Log Analytics, Azure Monitor Agent and Sentinel are not the agent tracing components.

For GPT-image edits, setting input_fidelity to high makes the model preserve fine details of the input image, such as product shape, logos and faces, so the output keeps the product identity. Simply sending the image and prompt uses low fidelity by default, groundedness detection is for text, and image models have no temperature control here.

OCR text appended to the prompt is third-party grounding content, so hidden instructions in it are an indirect prompt injection, which Prompt Shields for documents detects. Prompt Shields for user prompts targets direct jailbreaks typed by the user, image moderation checks harmful visuals, and protected material detects copyrighted text.

Advanced data parsing (layout-aware ingestion using Document Intelligence/Content Understanding layout) performs OCR on scanned PDFs, keeps tables and headings as structure-aware chunks and adds page-number metadata. Basic parsing with fixed-size or page-level chunking splits tables and loses structure.

Groundedness and relevance are AI-assisted quality evaluators (an LLM acts as the judge), and harmful content categories such as hate, violence, sexual and self-harm are risk and safety evaluators. NLP metrics such as BLEU or F1 and the similarity and fluency evaluators do not score groundedness or harm.

In Custom Vision you create a project, upload and tag sample images of good and faulty components, and then train. Classifying each component image as faulty or a specific fault type is a classification project. Initialize the training dataset is not a Custom Vision step, and object detection would only be needed to locate several faults inside one image.

The Azure OpenAI embeddings models (such as text-embedding-3-small/large) return numeric vectors that represent text, so you can compare documents with cosine similarity with minimal code. GPT-3.5 and GPT-4 generate text, and DALL-E generates images.

The Read OCR API in Azure Vision extracts printed and handwritten text from images and is designed for large asynchronous volumes, with minimal development. Image Analysis returns captions and tags, Azure Language needs text input, and Document Intelligence targets forms and structured documents rather than plain article text.

A private endpoint lets Azure apps reach Search1 over a private network while public access is disabled. Assigning the Search Index Data Reader role scoped to an index limits an app to querying that index, whereas query keys apply to all indexes on the service.

Azure Vision Image Analysis returns image tags and supports a language parameter that includes English, French and Spanish, so no training or translation is needed. Custom Vision needs training and returns your own labels, Content Safety classifies harm, and Custom Translator translates text.