Free AI-102 Exam Flashcards

A multi-service resource (kind CognitiveServices) exposes one endpoint and one key pair across Vision, Language, Speech, Translator, Document Intelligence, and Content Safety with consolidated billing. Single-service resources give a dedicated endpoint per capability so you can apply per-service RBAC, quotas, and network isolation, which is the production-grade choice.

Azure OpenAI always needs its own resource (kind OpenAI); it cannot live inside a multi-service CognitiveServices resource. If a scenario mixes GPT, DALL-E, or embeddings with Vision or Document Intelligence, you provision a dedicated Azure OpenAI resource plus a separate resource for the other services.

A Foundry hub holds shared compute, storage, key vault, and connections that many projects can reuse. A project is a workspace inside a hub that organizes the assets for one solution, so shared connections belong at the hub or resource scope to avoid duplicating credentials across projects.

Region choice is foundational because model availability and compliance capabilities vary by region. For European data residency and private connectivity, you must pick a supported European region and add private networking; application code cannot override the physical region where the service runs.

Provisioned throughput reserves capacity and gives predictable performance for steady, latency-sensitive production traffic. Standard deployment bills on actual consumption, so it is the better starting point for spiky prototypes and workloads with uncertain traffic.

Repeatable creation of Azure AI resources, private endpoints, and configuration across dev, test, and prod should use infrastructure as code in a CI/CD pipeline. IaC prevents configuration drift, makes environment promotion reviewable, and is far safer than manual portal configuration copied by hand.

Use separate Azure AI resources and project boundaries for development, test, and production to isolate quotas, credentials, and the blast radius of deployment changes. Sharing one production-like resource across all stages creates avoidable operational and security risk and makes quota issues hard to triage.

A system-assigned managed identity lets an Azure-hosted app obtain Microsoft Entra tokens without storing secrets in code or configuration, with access then granted via Azure RBAC. Keys in client-side JavaScript, hardcoded service-principal secrets, and SAS tokens all create leakage or rotation risks that managed identity is designed to avoid.

DefaultAzureCredential tries several authentication methods in a standard order, so the same code path can use local developer credentials on a workstation and managed identity after deployment to Azure. It reduces environment-specific branching when you use Azure SDK client libraries.

To restrict an Azure AI resource to private network clients only, create a private endpoint and disable public network access. IP filtering and CORS help with other concerns, but they do not by themselves close the public endpoint the way disabling public network access does.

A 429 response means the endpoint is throttling the caller; the correct client behavior is to retry with exponential backoff and then review quotas, request patterns, and concurrency. Hammering the endpoint harder or retrying immediately compounds the throttling rather than relieving it.

Standalone Azure AI Content Safety moderates arbitrary user content or images anywhere in a pipeline, including text that never touches a GPT model. Azure OpenAI's built-in content filters apply only to that deployment's prompts and completions, so user-generated-content moderation outside the model call must use Content Safety.

Content Safety lets you set separate severity thresholds for hate, sexual, violence, and self-harm, so a gaming chat can allow mild profanity while still auto-rejecting high-severity violence or self-harm. A single global block-all rule cannot match that risk profile and tends to over-block legitimate content.

Blocklists match known exact strings deterministically and update instantly with no training, which is ideal for org-specific terms and product names. Built-in harm categories generalize to phrasings you never listed but require classifier scores and severity thresholds, so blocklists are too brittle for fuzzy concepts.

Prompt Shields detect direct jailbreak attempts in user prompts and indirect attacks hidden inside retrieved documents, blocking attempts to steer the model away from its system instructions. Pair them with harm detection, retrieved-content validation, and least-privilege tool permissions for layered defense.

A model card summarizes what a model is intended for, where it may perform poorly, and what evaluation evidence exists, which makes it a core Responsible AI input for production approval. Without it, you risk deploying a model outside its supported use case and missing known limitations.

Azure OpenAI hosts the same OpenAI models inside Azure so you get Microsoft Entra identity, private networking, regional data residency, content filters, and Azure governance. Public OpenAI gives none of those enterprise controls, which is why regulated workloads use the Azure-hosted deployment.

For semantic retrieval, the same embedding model deployment must vectorize both document chunks at index time and user queries at query time so the two vectors live in the same space and can be compared. Text-to-speech, OCR, and content safety deployments solve different tasks and do not produce retrieval embeddings.

Fine-tuning is the right move when a stable behavior or style requirement persists despite strong prompting and you have quality labeled examples, not the default starting point. RAG and prompt engineering are cheaper and faster to iterate, so prefer them first when the gap is grounding or instruction-following.

Chat-model fine-tuning datasets are typically JSONL records of structured conversation examples, not raw PDFs or CSVs of embeddings. The supervised conversational format is what teaches the model the target tone or behavior pattern.

Retrieval-augmented generation fetches relevant content from your index at runtime and adds it to the prompt, which reduces reliance on what the model memorized during training. For internal policy or current manuals that the model was never trained on, RAG is the only way to get grounded answers.

Indexing a long PDF as one document makes retrieval miss the relevant passage because the search engine returns the whole unit. Chunking into smaller sections and indexing them separately lets Azure AI Search return the specific passage that matches the question and improves RAG answer quality.

Each chunk needs its embedding vector plus the actual chunk text and source metadata such as title, URL, or document ID. A vector alone is not enough to present grounded answers with usable citations, because the metadata is what lets the application show where the answer came from.

The system message is the place for durable, high-priority instructions that define the assistant's role, tone, and constraints across every turn. Putting that guidance only in user text or UI copy makes it easy to override or drift, because user messages are not the right place for stable behavior policy.

Prompt flow is for composing multi-step generative workflows such as retrieval, prompt templates, Python logic, and output parsing as reusable nodes that can be tested end to end and evaluated. The chat playground is for one-off prompt trials, not reproducible pipelines that ship to production.

A tool is a callable capability the agent can invoke, such as search, a function, or an external API, to retrieve data or perform an action beyond generating text from its prompt context alone. Tools expand what the agent can do; without them, the agent can only reason over what is already in its prompt.

Business rules that affect money, security, or compliance, such as payment validation, should stay in deterministic code or workflow steps, not be delegated to flexible LLM reasoning. Agentic flexibility is best applied around the language-heavy parts of the work where strict rules would be brittle.

A supervisor pattern separates responsibilities by routing work to specialized agents and then arbitrating their outputs, which makes complex workflows easier to maintain and evaluate than one overloaded agent. Single-agent designs are simpler and should be preferred when one tool-free response can handle every request.

Autonomous agents need explicit runtime controls such as max turns, tool-call limits, stop conditions, and approval gates for sensitive actions, because open-ended reasoning can loop and exhaust budget. Telemetry alone surfaces loops; it does not prevent them, so controls must be enforced in the runtime, not just monitored.

Image Analysis in Azure Vision returns Microsoft's prebuilt captions, tags, objects, and OCR for a still image with no training. Custom Vision is the choice when you have your own labeled images and need domain-specific classes that the prebuilt model does not know about.

Classification predicts what is in the image; object detection also predicts where it is, which is why it requires bounding-box labeling during training. Use detection when location matters, such as defective parts in a manufacturing image, and classification when only the presence of a class matters.

Exportable Custom Vision models for mobile or edge devices require a compact domain, which is optimized for constrained environments. Standard domains do not support export, so pick the compact domain at training time if you need to ship the model offline.

Line crossing emits events when someone crosses a configured boundary, which is the direct fit for entrance or exit counts. Zone dwell time uses a polygon and measures how long each person remains inside it, which is the right operation for queue monitoring or service-time analysis.

Video Indexer extracts multimodal insights from recorded video, including transcripts, speakers, keywords, scenes, and timestamps, in one service. The other Vision services focus on still images or custom training, not multimodal video insights across audio and visual tracks.

Speech-to-text transcribes spoken audio into text for downstream analytics, captions, or further NLP. Once audio is transcribed, other services can analyze the resulting text for sentiment, key phrases, or PII; speech-to-text is not itself a translation or analysis service.

The say-as element tells the speech engine how to interpret text such as dates, times, phone numbers, or digits, for example reading 03/08/2026 as a date instead of individual characters. It is the standard way to fix awkward readings without retraining the voice model.

When the baseline speech model misrecognizes company brand names, drug names, or industry acronyms despite good audio, train a Custom Speech model with domain audio and transcripts. Custom Speech adapts recognition to your vocabulary and acoustic environment instead of forcing you to add post-hoc correction logic.

Sentiment analysis classifies text as positive, negative, mixed, or neutral and is the right fit when the input is already text and the goal is opinion or satisfaction. It does not translate text, extract entities, or detect intent; those are separate Language service features.

PII detection finds sensitive entities such as Social Security numbers, credit card numbers, and other identification or financial data in text, and supports redaction so downstream users see masked content. It is the right tool for compliance workflows before analysts review chat logs or documents.

CLU maps user utterances to a fixed, trained set of intents and entities, which is the right fit when a bot must decide what action the user wants and capture parameters like an order number. It is deterministic and bounded, unlike Azure OpenAI, which generates free-form text and can hallucinate.

Custom question answering builds a knowledge base from FAQ pages and PDF manuals and returns the best answer to user questions without hand-authoring every response. CLU is different: question answering matches a query to source content, while CLU classifies an utterance against trained intents.

Custom Translation is the right fix when the generic Translator mishandles approved domain terms such as product jargon, part names, or safety phrases. Training with bilingual parallel examples steers output toward the approved vocabulary instead of letting the generic model substitute wrong words.

An indexer pulls data from supported sources such as Azure Blob Storage and populates a search index automatically on a schedule with minimal custom ETL. Attach a skillset to enrich content during indexing, for example running OCR and key phrase extraction on scanned PDFs before they become searchable.

Knowledge store persists enrichment outputs from a skillset into structured projections in Azure Storage so BI tools can consume them outside the search index. It complements the index by retaining intermediate or enriched data for downstream analytics, not just for query-time retrieval.

Hybrid search combines lexical and vector retrieval in one request, which is a strong default for production search because it blends exact term matching with semantic similarity. Semantic reranking on top improves final ordering and helps when users mistype or use different words for the same concept.

Nearest-neighbor retrieval over embeddings requires vector-enabled fields in the index and vector queries at query time, which is how the engine compares the query embedding to stored chunk embeddings. Semantic captions or analyzers alone do not enable vector retrieval.

A vectorizer lets Azure AI Search convert a plain-text user query into an embedding at query time, so the client does not have to call a separate embedding model first. Pair it with an indexing pipeline that generates embeddings during ingestion for end-to-end built-in vectorization.

The prebuilt invoice model already understands common invoice structure and fields such as vendor name, total, and line items, so it is the fastest path for standard invoice extraction. Build a custom model only when your forms have fields or layouts the prebuilt model does not cover.

A composed model groups multiple trained custom extraction models behind one model ID; the service decides which submodel fits each submitted document and routes extraction accordingly. That simplifies application logic when several business form types must share one endpoint.

Content Understanding in Foundry Tools extracts structured fields from documents, images, audio, and video into one user-defined schema. It is broader than document-only Document Intelligence because one pipeline can process call recordings, screenshots, PDFs, and videos against the same business schema.