Free AI-103 Exam Flashcards

Microsoft Foundry is the primary environment for building, managing, evaluating, and deploying Azure AI apps and agents.

Choose the model based on task requirements such as reasoning depth, latency, cost, modality, context needs, and deployment constraints.

Use a small language model when the task is narrow, latency or cost matters, and deep reasoning is not required.

Use a multimodal model when the app must reason over more than one input type, such as text with images, audio, or video context.

Keyless authentication uses Microsoft Entra ID and managed identity instead of embedded API keys. It improves secret management and access control.

Managed identity lets Azure-hosted compute access services through assigned identity and RBAC, avoiding hard-coded credentials.

Private networking reduces public exposure for services and data paths. It is used when AI resources, search indexes, storage, or apps must stay inside controlled network boundaries.

They control how much model capacity an app can use over time. Monitor them to avoid throttling, failed requests, or unexpected scaling limits.

Track model performance, latency, cost, safety events, drift, search relevance, grounding quality, and app errors.

It is the use of evaluators, safety checks, trace logs, provenance, and oversight workflows to measure and govern AI behavior.

Content filters classify and block or flag unsafe inputs and outputs. They are guardrails, not a replacement for app-level validation and monitoring.

Prompt injection is an attempt to override instructions or misuse tools through user input, retrieved content, or embedded data. Defenses include isolation, filtering, tool limits, and validation.

RAG retrieves relevant grounding content and adds it to the model context so generated answers can be based on approved sources.

Grounding connects model responses to trusted data, retrieved documents, or tool results. It helps reduce unsupported answers.

Function calling lets a model request a structured tool or API call. The app executes the call and returns results for the model to use.

Structured output asks the model to return data in a defined schema such as JSON. It helps apps validate, parse, and route model responses.

Prompt engineering sets role, task, constraints, examples, output format, and safety instructions so the model produces useful and predictable responses.

Temperature affects response variability. Lower values are better for consistent outputs; higher values can make responses more creative but less predictable.

Evaluation measures whether a model or app meets goals for relevance, quality, safety, groundedness, and task success before and after deployment.

Tracing records the steps in an AI workflow, such as prompts, tool calls, retrieval, latency, and outputs. It helps debug and audit behavior.

Token analytics helps track usage, cost, prompt size, response size, and performance trends across AI app calls.

An agent role defines what the agent is responsible for, how it should behave, and what boundaries apply to its responses and actions.

A tool schema describes a callable tool's name, purpose, inputs, and expected structure so the model can request tool use correctly.

Conversation memory preserves relevant prior context for an agent. It should be scoped carefully so old or sensitive context does not pollute future actions.

Multi-agent orchestration coordinates specialized agents or steps to complete a broader workflow, often with routing, handoffs, and approval gates.

Require approval before high-impact, irreversible, costly, or sensitive actions such as sending external messages, changing records, or executing privileged tools.

Visual question answering uses a model to answer questions grounded in image or video content, such as identifying visible defects or comparing visual elements.

Image captioning generates short or detailed text descriptions of visual content. It can support search, accessibility, and review workflows.

Alt text generation creates concise descriptions of images for accessibility. It should describe meaningful content without adding unsupported details.

Inpainting edits a selected region of an image while preserving the surrounding context. It is used for controlled image modification.

Text embedded in images can carry indirect prompt injection instructions. Treat visual text as untrusted input before using it to guide an agent.

Object detection identifies and locates objects or regions within an image or video frame. It is useful when the app needs more than a general caption.

Text analysis extracts meaning from text, such as entities, topics, sentiment, summaries, sensitive content, or structured data.

Entity extraction identifies important items in text, such as people, organizations, dates, products, locations, or custom business terms.

Sentiment analysis estimates the emotional tone of text, such as positive, negative, neutral, or mixed. It is useful for feedback and support scenarios.

Sensitive content detection finds content such as personal data, unsafe material, or policy-relevant text so the app can redact, block, route, or review it.

Speech-to-text converts audio into text so apps and agents can transcribe, analyze, summarize, or respond to spoken input.

Text-to-speech converts text responses into spoken audio. It is used for voice agents, accessibility, IVR, and conversational apps.

Translation lets an app accept or produce content across languages. It may use dedicated translation services or model-based translation flows.

Domain customization adapts prompts, schemas, examples, or models to a specific business vocabulary, format, or compliance requirement.

Information extraction turns documents or multimodal content into usable data, such as fields, summaries, markdown, tables, entities, or grounded chunks.

OCR converts text in images, scans, or documents into machine-readable text for search, extraction, review, or RAG pipelines.

Azure AI Search indexes content and retrieves relevant chunks for grounding. It can support keyword, vector, hybrid, and semantic search patterns.

Vector search finds content by semantic similarity using embeddings. It is useful when exact keyword matches are not enough.

Hybrid search combines keyword and vector retrieval. It can improve relevance when queries need both exact terms and semantic similarity.

Semantic ranking reorders search results based on meaning and relevance. It is commonly layered on top of retrieved candidates.

Indexing prepares content for search by extracting, chunking, enriching, embedding, and storing fields that retrieval can query.

Enrichment adds derived information such as OCR text, entities, captions, translations, key phrases, or embeddings before content is indexed.

Content Understanding extracts structured or markdown outputs from multimodal content such as documents, images, audio, or video for downstream reasoning.

Use trusted sources, access controls, security trimming, prompt-injection defenses, citation/provenance data, and monitoring for grounding quality.