AI-103 cram sheet
Every flashcard fact for AI-103, condensed into one scannable page and grouped by exam domain. Built for the last hour before the exam — and for printing: use your browser's print or save-as-PDF and the navigation, sidebar, and controls drop away.
D1 · Plan & manage AI 25–30%
| What does a 429 with 'service is temporarily unable to process your request' indicate, and the fix? | Transient capacity throttling (not quota). Retry with exponential backoff honoring retry-after-ms; consider PTU for consistency. |
|---|---|
| What three things define Global Batch versus Global Standard? | About 50% lower token cost, a 24-hour target turnaround with no real-time SLA (jobs that run longer keep going until they finish or you cancel), and a separate enqueued token quota so online workloads aren't disrupted. |
| What are the four content filter harm categories and four severity levels? | Categories: hate/fairness, sexual, violence, self-harm. Severity levels: safe, low, medium, high (safe is annotated only, not filtered). |
| Name the SKU codes for the Foundry deployment types. | GlobalStandard, GlobalProvisionedManaged, GlobalBatch, DataZoneStandard, DataZoneProvisionedManaged, DataZoneBatch, Standard (regional), ProvisionedManaged (regional provisioned), DeveloperTier. |
| Managed compute vs serverless deployment: how does billing differ? | Managed compute bills per VM core-hour on dedicated VMs you manage; serverless bills per input/output token for a Microsoft-hosted API. |
| What is the Developer (DeveloperTier) deployment type for, and what are its limits? | Cost-efficient evaluation of fine-tuned models only. No SLA, no data residency guarantee, and a fixed 24-hour lifetime after which the deployment is deleted automatically. |
| Order the Foundry built-in roles from least to most privileged, and say what each adds. | Foundry Agent Consumer (call agent endpoints only) → Foundry User (build/develop in a project, reader access) → Foundry Project Manager (adds publishing agents and assigning Foundry User) → Foundry Account Owner (create accounts/projects, manage models, but not build) → Foundry Owner (everything, incl. build). |
| What does the groundedness detection filter do and what is its constraint? | Flags whether an LLM response is grounded in the provided source materials; available only in streaming scenarios with document embedding/formatting. |
| Which project type is required to deploy managed-compute models like Hugging Face models? | A hub-based project (classic). Foundry projects do not support managed-compute deployments. |
| How much indexer execution history does Azure AI Search keep, and what statuses can a run have? | Up to the 50 most recent runs, most recent first, via Get Indexer Status or the portal's Execution details. Per-run statuses: running, success, transientFailure, persistentFailure, plus a separate Reset entry when change-tracking state is cleared. The top-level indexer status "running" only means it is available to run. |
| Where is inference data processed for Global vs DataZone vs Standard deployment types? | Global: any Azure region. DataZone: only within the Microsoft-specified data zone (US, EU, or APAC). Standard/Regional: the deployment region. Data at rest always stays in the designated Azure geography regardless of type. |
| How do you authenticate to Foundry endpoints without storing keys? | Use Microsoft Entra ID with a managed identity and Azure RBAC (keyless credentials). |
| What are the managed virtual network isolation modes for a Foundry hub? | No isolation, Allow internet outbound, and Allow only approved outbound (FQDN/service-tag/private-endpoint outbound rules). |
| Why can an indexer report Success while itemsFailed is greater than zero? | A run succeeds as long as document errors stay below the indexer's Max failed items setting; those documents are skipped. The run is marked Failed only once errors exceed the threshold. Warnings (for example, image-only PDFs with no extractable text) are common and don't fail a run. |
| What are the two top-level categories in the Foundry model catalog? | Models sold by Azure (Azure Direct models, Microsoft-hosted with enterprise SLA/support) and Models from partners and community (third-party/community, supported by providers). |
| What is required to turn Azure OpenAI content filters off or to annotate-only? | Approval via the Modified Content Filters limited access review; otherwise these options are unavailable. |
| What four things can you configure on a Foundry agent's Monitor settings panel? | Continuous evaluation (evaluators run on a sampled share of production responses; default limit 100 runs/hour via max_hourly_runs), scheduled evaluations (fixed test dataset on a schedule — the drift signal), red team scans (scheduled adversarial testing), and alerts on latency, token usage, evaluation scores, or red team findings. All require Application Insights connected to the project. |
| What do Prompt Shields protect against? | User prompt attacks (jailbreaks) and indirect/cross-domain prompt injection attacks embedded in documents the system processes. |
| When should you choose Provisioned Throughput (PTU) over Standard? | For latency-sensitive, high-volume production workloads needing dedicated capacity and guaranteed, predictable throughput instead of a shared pool. |
| Which least-privilege role lets you view quota usage, and at what scope? | Cognitive Services Usages Reader, assigned at the subscription level. |
| On the Foundry Quota page, what does the Scope column tell you? | Global or Data Zone = subscription-level quota management: all deployments of that model and version share one pool (per subscription globally, or per data zone). An actual region name = per-region quota for that subscription and model. Spreading Global Standard deployments across regions does not add capacity when scope is Global. |
| In what unit is Azure OpenAI quota allocated, and at what granularity? | Tokens-Per-Minute (TPM), allocated per subscription, per region, per model, per deployment type. RPM is derived proportionally from TPM. |
| Why should you not assign Azure AI Developer or Cognitive Services roles for Foundry project work? | Despite its name, Azure AI Developer is scoped to Azure Machine Learning workspaces and Foundry hubs, not Foundry projects or hosted agents. Roles beginning with "Cognitive Services" target AI Services resources directly. Use Foundry User or Foundry Owner instead. (Foundry User/Owner/Account Owner/Project Manager were formerly Azure AI User/Owner/Account Owner/Project Manager — the role IDs are unchanged.) |
| How are Azure AI Search search units calculated, and what does each dimension buy you? | SU = replicas × partitions, and SU is the billing unit. Replicas = query throughput and availability (2+ for the query SLA, 3+ for read-write). Partitions = storage and indexing throughput. Partition count has no effect on SLA. |
| When are you billed for the semantic ranker? | When a request sets queryType=semantic AND the search string is non-empty. search=* with queryType=semantic is not billed. A free monthly allowance applies by default before the standard pay-as-you-go plan. |
| What exactly does the semantic ranker rerank, and how is the score expressed? | It reranks only the top 50 results from the BM25 or RRF pass — it cannot rerun the query over the corpus. Each document gets @search.rerankerScore from 4 (highly relevant) down to 0 (irrelevant). Summarization takes up to 2,000 tokens per document, so semantic configuration field order matters. |
| How is spillover configured and identified? | Set spilloverDeploymentName on the provisioned deployment (a standard deployment of the same model/version in the same resource), or send x-ms-spillover-deployment per request. Triggers on 429 (PTUs exhausted), 400 (long context), 500, 503. Spilled responses carry x-ms-spillover-from-deployment and x-ms-spillover-error; in metrics, split on IsSpillover. |
D2 · Generative AI & agents 30–35%
| What three components define every Foundry agent? | Model (reasoning), Instructions (goals/constraints/behavior), and Tools (data/actions). |
|---|---|
| What three core components combine to form every Foundry agent? | Model (from the catalog), Instructions (goals/constraints/behavior), and Tools (data or actions). |
| Name agent-specific evaluators. | Intent resolution, task adherence (task completion), and tool call accuracy. |
| How can an agent securely access external resources/MCP servers without shared credentials? | Give it a dedicated Microsoft Entra agent identity (managed identity) with RBAC; OAuth On-Behalf-Of (OBO) passthrough is also supported. |
| What is the AI Red Teaming Agent used for? | Simulating adversarial attacks (via Microsoft PyRIT) to find safety/security vulnerabilities before deployment, ideally with human-in-the-loop. |
| What does the Azure AI Search agent tool do? | Grounds an agent with data from an existing Azure AI Search index via a project connection. |
| How do you run a Foundry agent asynchronously for long tasks? | Set background=True on responses.create(), then poll responses.retrieve(response.id) while status is queued or in_progress. Streaming (stream=True) is the alternative when you want incremental output instead. |
| What is chain-of-thought prompting and when does it apply? | Instructing the model to work step-by-step and present its reasoning to reduce inaccuracy; applies only to non-reasoning models. |
| Why ask the model to cite sources, and which citation style is stronger? | Citations make statements more grounded (the model must make two errors to be wrong); inline citations are stronger mitigations than end-of-content citations. |
| What does the Code Interpreter tool provide? | Writing and running Python in a sandboxed environment for data analysis, math, and chart generation. |
| Two ways to carry multi-turn context in the Foundry Responses API? | Create a durable conversation object and pass its ID on each request, or omit the conversation and pass previous_response_id from the prior response. Conversations persist across sessions and store items (messages, tool calls, tool outputs). |
| Handoff vs agent-as-tools — what's the difference? | Handoff passes control explicitly; the receiving agent takes full ownership and has the whole conversation context, with no central orchestrator (mesh topology). Agent-as-tools keeps the primary agent responsible — it delegates a subtask and control returns when the subtask finishes. |
| What is few-shot learning? | Including example input/output pairs in the prompt to condition the model for the current request only—no retraining. |
| When do you use the File Search tool? | To ground an agent in uploaded files/proprietary documents using vector search over a vector store. |
| What is a Foundry IQ knowledge base, and what powers retrieval? | A shareable, permission-aware multi-source knowledge base (knowledge sources + retrieval params); agentic retrieval over Azure AI Search decomposes, parallelizes, and reranks queries. |
| How do you return a tool result to a Foundry agent? | The agent emits an output item of type function_call with a call_id and arguments. Your app runs the function, then creates a follow-up response whose input contains a function_call_output item with the same call_id and the output payload. |
| After an agent emits a function_call, what must your app do? | Execute the function and submit a function_call_output item with the matching call_id so the agent can finish its answer. |
| What do the Groundedness and Relevance evaluators measure? | Groundedness: is the response supported by the provided context? Relevance: does it address the user's query? Both are RAG-specific metrics. |
| What is a Hosted agent? | A code-based agent (Agent Framework, LangGraph, OpenAI Agents SDK, etc.) packaged as a container/zip; Foundry runs it with a managed endpoint, autoscaling, and a dedicated Entra identity. |
| json_object vs json_schema response_format? | json_object (JSON mode) guarantees valid JSON but not schema adherence. json_schema with strict: true forces the model to follow your supplied JSON Schema. Structured outputs also require parallel_tool_calls=false. |
| When is the MCP tool the best choice? | For tools hosted on an MCP server endpoint that are shared across multiple agents or maintained by another team. |
| The three long-term memory types in Foundry Agent Service? | User profile memory (durable preferences — retrieve early in a conversation), chat summary memory (distilled prior topics — retrieve per turn), procedural memory (reusable how-to routines). Phases: extraction, consolidation, retrieval. |
| What does the OpenAPI tool connect an agent to? | External HTTP APIs described by an OpenAPI 3.0 or 3.1 spec, with anonymous, API key, or managed identity auth. |
| What is a Prompt agent in Foundry Agent Service? | An agent defined entirely by configuration (instructions, model, tools) that Foundry runs for you—no app code, compute, or containers to manage. |
| What must be true for a prompt to hit the cache? | At least 1,024 tokens, with the first 1,024 tokens identical; after that, hits occur every 128 additional identical tokens. One character of drift in the prefix gives cached_tokens = 0. Put static instructions first, variable content last. |
| Prompt agent vs Hosted agent: key difference? | Prompt agents are config-defined and fully Foundry-run (no code/compute to manage); Hosted agents run your container code with Foundry-managed endpoint/scaling. |
| How does RAG reduce fabrication? | It retrieves relevant grounding data and injects it into the prompt so the model answers from authoritative source content instead of parametric memory. |
| Name key RAG quality evaluators in Foundry. | Groundedness and relevance (plus general-purpose coherence and fluency); safety evaluators like hate/unfairness, violence, self-harm. |
| Which token-cap parameter do reasoning models take? | max_completion_tokens on the Chat Completions API and max_output_tokens on the Responses API. Legacy max_tokens is not accepted. Temperature, top_p, presence_penalty, and logprobs are unsupported on reasoning models. |
| Which usage fields matter for reasoning-model cost analysis? | completion_tokens_details.reasoning_tokens holds the hidden thinking tokens; prompt_tokens_details.cached_tokens shows prompt-cache hits. reasoning_effort (low/medium/high) drives how many reasoning tokens are produced. |
| What is recency bias and how do you mitigate it? | Text near the end of a prompt has more influence; mitigate by repeating key instructions at the end. |
| What is the Responses API in Foundry Agent Service? | The single entry point giving any framework/process access to Foundry models plus platform tools (file search, code interpreter, web search, MCP, memory). |
| What is the Responses API in Foundry? | The single model-and-tools entry point behind every agent type; callable from your own process to get Foundry models and platform tools without moving your code. |
| Where do you set the content filter and TPM limit for model router? | On the model router deployment itself — the settings apply to all underlying chat models. Don't deploy the supported OpenAI models separately or configure filters/rate limits per model. Claude models are the exception and must be deployed first. |
| Model router routing modes and when to use each? | Balanced (default) for most workloads; Quality for critical tasks like legal or medical review and complex reasoning; Cost for high-volume, budget-sensitive work such as classification. Mode changes take up to five minutes. |
| How long after creation do agent runs expire, and what should slow operations do? | 10 minutes; for slow operations, return a status immediately and poll separately. |
| Your function tool calls an API that can take 15 minutes. What's the recommended pattern? | Runs expire 10 minutes after creation and the limit covers total elapsed time, so return a status immediately from the tool and poll for the real result separately rather than blocking inside the call. |
| What changes when you set store=False on a Responses call? | The service doesn't persist the response server-side, so previous_response_id is unavailable and your app must carry context forward by passing previous output items as input. Used for zero-data-retention scenarios. |
| Three rules for a strict structured-outputs JSON Schema? | additionalProperties: false on every object; every property listed in required (emulate optional with a ["string","null"] union); max 100 object properties and 5 levels of nesting. Root cannot be anyOf. |
| What do structured inputs enable for agent tools? | Overriding tool values (e.g., vector_store_ids, MCP server_url/headers) at runtime without creating a new agent version. |
| How does temperature affect output, and what about top_p? | Low temperature (e.g., 0.2) = focused/concrete; high = random/creative. Adjust only one of temperature or top_p at a time, not both. |
| What does the Tool Call Accuracy evaluator measure? | Whether the agent made the right tool calls with correct parameters; a process evaluator output as Pass/Fail from a 1-5 threshold. |
| What is a Foundry Toolbox? | A curated bundle of tools configured once and exposed as a single MCP-compatible endpoint (with versioning and central auth) for reuse across agents. |
| What standard underpins Foundry distributed tracing and where do traces go? | OpenTelemetry, integrated with Azure Monitor Application Insights; captures LLM calls, tool invocations, and agent decisions. |
| What standard is Foundry distributed tracing built on, and where do traces land? | OpenTelemetry, integrated with Azure Monitor Application Insights. |
| What is Microsoft's recommended way to add web grounding to an agent? | The Web search built-in tool, which retrieves real-time public web info and returns answers with inline citations. |
| What risk do Foundry content filters help mitigate for agents ingesting untrusted content? | Cross-prompt injection attacks (XPIA), where hidden instructions in untrusted data try to hijack the agent. |
D3 · Computer vision 10–15%
| Image Analysis 4.0: Caption vs Dense Captions? | Caption returns one human-readable sentence for the whole image — the natural alt-text string. Dense Captions returns a whole-image sentence plus up to 10 region descriptions, each with bounding box coordinates, which is what an extended accessibility description needs. Both are English-only and only available from captioning-enabled regions. |
|---|---|
| What error code and status indicate an image generation request was blocked by the safety system? | Operation status Failed with error.code = contentFilter. The prompt or the generated image may be the cause, indicated by the message text. |
| What severity values does the Azure AI Content Safety image model return? | Only the trimmed scale 0, 2, 4 and 6 (safe / low / medium / high) across the four harm categories Hate, Sexual, Violence and SelfHarm. Text and multimodal image-with-text models support the full 0-7 scale and can be asked to return the trimmed scale instead. |
| What is Azure Content Understanding? | A Foundry tool that uses generative AI to process documents, images, video, and audio into user-defined structured output (JSON schema or Markdown). |
| What does an analyzer configure in Content Understanding? | Content-extraction settings, the field-extraction schema, and model deployments. It applies these consistently to all incoming data. |
| Content Understanding image analyzer input limits? | .jpg/.jpeg/.jpe, .png, .bmp, .heif, .heic; up to 200 MB; resolution from 50 x 50 up to 10,000 x 10,000 pixels. If the analysis is driven mainly by text in the image, use a document field extraction schema instead. |
| What does enableSegment control on a Content Understanding video analyzer? | false = whole-video mode: the entire file is one segment, good for compliance sweeps and full-length summaries. true = custom segmentation: you describe the slicing logic in natural language via contentCategories and the generative model creates matching segments (chapters, news stories). Setting segmentation consumes model tokens even with no fields defined. |
| Name the three Content Understanding field-extraction methods. | Extract (values as they appear, documents only), Classify (assign a category, can route to another analyzer), and Generate (freely create values like summaries/scene descriptions). |
| How do you stop Image Analysis 4.0 captions from using gendered terms? | Set the optional request parameter gender-neutral-caption to true on the Analyze Image call. By default captions use "man", "woman", "boy" and "girl"; with the flag on, they are replaced with "person". |
| What do the n and quality parameters control in gpt-image generation? | n = number of images per request (1-10); quality = low/medium/high, where low is optimized for latency-sensitive use cases. |
| What does grounding provide in Content Understanding, and how is it enabled? | It traces each extracted value back to the source region for verification. Enabled with the estimateFieldSourceAndConfidence setting. |
| Which Image Analysis features are only in version 3.2, not 4.0? | Brands, image categorization, faces, image type, domain models (celebrities/landmarks), color scheme, and adult-content moderation. |
| Image Analysis 4.0: Caption vs Dense captions? | Caption = one human-readable sentence for the whole image. Dense captions = detailed captions for each individual object plus bounding boxes. |
| Which features are exclusive to Image Analysis version 4.0? | Synchronous Read (OCR), People detection, plus improved Captions, Dense captions, Object detection, Tags, and Smart crop. |
| Image Analysis 4.0 input requirements? | JPEG, PNG, GIF, BMP, WEBP, ICO, TIFF or MPO; under 20 MB; dimensions greater than 50 x 50 and less than 16,000 x 16,000 pixels. (Version 3.2 is stricter: JPEG/PNG/GIF/BMP and under 4 MB.) |
| What does inpainting on the gpt-image edits endpoint require besides the image and prompt? | A mask the same size as the input image, marking the region to regenerate. Only the masked area is changed per the prompt. |
| Rules for a mask on the Foundry images/edits endpoint? | The mask is a PNG with an alpha channel; fully transparent pixels (alpha 0) mark the region the model should regenerate. It must be exactly the same dimensions as the input image. The image being edited must be PNG or JPG and under 50 MB. input_fidelity raises preservation of the original style and features (not available on gpt-image-1-mini). |
| Which model series replaced the retired dall-e-3 for Azure OpenAI image generation? | The gpt-image series (e.g., gpt-image-1, gpt-image-1.5, gpt-image-2). dall-e-3 was retired and its deployments are non-functional. |
| In what format does the gpt-image series return generated images? | Base64-encoded image data in the b64_json field of the response (no URL option). |
| What do Image Analysis 4.0 multimodal embeddings enable? | Vectorizing images and text into the same space so you can search images with natural-language text by semantic closeness, without tags. |
| Content Understanding pro mode: what does it add and what does it give up? | Adds multi-step reasoning, multiple input documents per request, and reference data supplied at analyzer creation — for validation scenarios such as invoice vs contract. Gives up confidence scores, grounding, and extract fields, and it currently supports documents only (.pdf, .tiff, images; up to 100 MB and 150 pages). |
| How are filtered/harmful results surfaced in a Content Understanding analyze response, and where is filtering tuned? | As a content_filters array in the response. Filtering is tuned in the Guardrails instance on the model deployment (adjust thresholds or switch to annotate mode). |
| How do you split a video into scenes for separate processing in Content Understanding? | Use segmentation via the enableSegment property in the analyzer schema to divide documents or videos into logical sections. |
| Key Sora 2 generation controls in Foundry? | prompt (required); optional model, size (portrait 720x1280 default, landscape 1280x720), seconds (4/8/12, default 4), input_reference (one image anchoring the first frame, jpeg/png/webp, must match the size exactly) and remix of a completed video's ID for targeted edits. Generation is asynchronous: create the job, poll status, then download the MP4. |
| Content Understanding video: analyzeBinary vs analyze? | analyzeBinary uploads the file in the request body and is capped at 200 MB and 30 minutes (this is what the Foundry and Studio UX use). analyze references the file by URL from Blob Storage and allows up to 4 GB and two hours. |
| What is the effect of the 'low resolution' detail setting on vision-enabled chat models? | Faster responses and fewer input tokens, but reduced accuracy of object and text recognition in the image. |
D4 · Text analysis 10–15%
| When do you use CLU instead of custom text classification? | CLU predicts the intent of a conversational utterance and extracts entities from it (intent + slots) for bots and command-and-control apps. Custom text classification assigns classes to whole documents and extracts nothing. CLU only understands the text — the client app performs the action. |
|---|---|
| Name the conversation summarization aspects and what each returns. | issue = the customer's problem; resolution = what was tried; recap = the whole conversation as one paragraph; chapterTitle = a title per topic segment; narrative = a summary per topic segment (chapterTitle and narrative are normally used together). |
| Single label vs multi label custom text classification — when do you pick each? | Single label assigns exactly one class per document; use it when classes are mutually exclusive and clearly separable. Multi label allows several classes per document; use it when a document can legitimately belong to more than one category. |
| When should you train a custom speech model? | To improve recognition of domain-specific vocabulary (train with relevant text) or accuracy for specific audio conditions (train with audio plus reference transcriptions). |
| When do you use Custom Translator vs. Adaptive custom translation? | Custom Translator builds domain/industry-tuned models. Adaptive custom translation uploads 5-10,000 prealigned segment pairs to build a bilingual index that adapts a supported LLM in minutes via the Translator APIs. |
| What gates the use of custom (neural) voice in Azure Speech? | Custom voice (professional voice fine-tuning / personal voice) requires approved limited access for responsible use, plus a handful of audio files and transcriptions to get started. |
| What does diarization do, and how many speakers can it identify? | It distinguishes and separates different speakers in a recording; the Speech service can identify up to 35 speakers (an error is thrown beyond that). |
| Synchronous vs. asynchronous document translation in Azure Translator? | Synchronous translates a single file and returns it directly (no Blob storage). Asynchronous (batch) translates many/complex files preserving structure and requires a Blob storage account. |
| What are EmbeddedSpeechConfig and HybridSpeechConfig used for? | EmbeddedSpeechConfig runs speech to text and text to speech entirely on-device from downloaded models and voices, with no Foundry resource — for intermittent or absent connectivity. HybridSpeechConfig uses the cloud by default and falls back to embedded when the connection fails. Embedded speech is limited access and is not supported by the Python or Go SDKs for hybrid. |
| How do the outputs of extractive and abstractive summarization differ? | Extractive returns salient sentences copied verbatim from the source, each with a rank score and positional information (start offset and length). Abstractive generates new, concise sentences that are not verbatim, each labelled with the contextual input range that produced it. |
| What does language detection return per document, and how do you disambiguate ambiguous text? | It returns the main language name, its ISO 639-1 code, a confidence score, and a script name plus ISO 15924 script code. To disambiguate words common to several languages, pass an ISO 3166-1 alpha-2 country/region hint. |
| Prebuilt NER vs. custom NER in Azure AI Language? | Prebuilt NER categorizes entities from a preset list (people, places, organizations, quantities) with no training. Custom NER is trained on labeled examples to recognize specialized, domain-specific entities. |
| What does opinion mining (aspect-based sentiment analysis) add over basic sentiment analysis? | It links sentiment to specific aspects/targets in the text (e.g., product or service attributes) along with their assessments, giving more granular opinion information. |
| What is the lightweight alternative to custom speech for boosting recognition of a few terms? | Phrase lists - supply specific words/phrases likely to be spoken (domain terms, proper nouns, uncommon words) to guide the model without training a custom model. |
| Which PII feature type is optimized for turn-based chat and transcripts? | Conversation PII - processes multi-turn exchanges asynchronously using conversation context and turn boundaries (e.g., contact-center transcripts). |
| Which PII feature type redacts native .pdf/.docx/.txt files while preserving structure? | Document-based PII - asynchronous, storage-based pipeline that preserves document structure and also produces machine-readable JSON metadata. |
| Which PII feature type processes raw text strings synchronously? | Text PII - optimized for synchronous, string-based input such as messages, prompts, logs, and ticket fields, returning detection and redaction results. |
| At what levels does Azure AI Language sentiment analysis evaluate text, and what scores does it return? | At both the document and sentence level; it returns positive, neutral, and negative confidence scores (0-1) for the document and each sentence, choosing the label with the highest score. |
| What does multi-lingual speech translation enable that standard speech translation doesn't? | No specified input language, language switching within the same session, and a transcription output in the target language - without restarting the session. |
| What is SSML used for in text to speech? | Speech Synthesis Markup Language: an XML-based markup to customize TTS output - adjust pitch, rate, volume, add pauses, improve pronunciation, define lexicons, and switch speaking styles or voices. |
| What two rules must every object in a strict structured-outputs schema follow? | Set additionalProperties: false, and list every field in required. Optional fields are emulated with a nullable union such as "type": ["string", "null"]. Keywords like pattern, format, minItems, and maxItems are unsupported; the root cannot be anyOf. |
| What does structured outputs give you that JSON mode does not? | JSON mode (response_format json_object) guarantees only that the reply is valid JSON. Structured outputs (response_format json_schema with strict: true) additionally guarantee the reply conforms to your JSON Schema — all required fields present, no extras. |
| Name the speech to text modes and their best use. | Real-time (live audio with intermediate results), Fast (synchronous, faster-than-real-time for prerecorded files), and Batch (asynchronous bulk processing). Custom speech improves domain accuracy. |
| What four things does Text analytics for health do in a single API call, and what special output format does it support? | Named entity recognition, relation extraction, entity linking to UMLS, and assertion detection (certainty, conditionality, association, temporality). It can also return results in FHIR structure for EHR integration. It is not for clinical decision making. |
| Does Azure Translator text translation require Blob storage? | No. Text translation runs real-time translation between supported languages via REST/SDK with no Blob storage. Asynchronous batch document translation is what requires source/target Blob containers. |
| How do profanityAction and profanityMarker interact in Translator v3? | profanityAction is NoAction (default, profanity passes through), Marked, or Deleted (word removed). When Marked, profanityMarker chooses Asterisk (default, replaces with ***) or Tag (wraps the word in <profanity>...</profanity> for post-processing). |
D5 · Information extraction 10–15%
| What is AI enrichment in Azure AI Search built on? | Skillsets composed of built-in skills (OCR, entity recognition, key phrase, translation, embeddings) and custom skills, run as part of an indexer pipeline. |
|---|---|
| What is Azure AI Content Understanding? | A multimodal service with customizable analyzers that turn documents, images, audio, and video into structured, strongly-typed, agent/RAG-ready output (structured or Markdown). |
| How do you enable confidence and grounding in Content Understanding field extraction? | Opt in by setting estimateFieldSourceAndConfidence = true (or per-field estimateSourceAndConfidence = true) to return source location and confidence scores. |
| How do you add proprietary processing to an enrichment pipeline? | Create a custom skill that wraps your external code behind the custom skill web interface, typically hosted as an Azure Function or Web API. |
| What are the two deletion detection policies for Azure Storage data sources? | NativeBlobSoftDeleteDeletionDetectionPolicy (Blob Storage only, needs blob soft delete enabled and blob versioning off) and SoftDeleteColumnDeletionDetectionPolicy with `softDeleteColumnName`/`softDeleteMarkerValue` over custom metadata. Either must be in place from the very first indexer run. |
| What is document cracking? | The first enrichment step where the indexer opens each source document and extracts its embedded text and images before skills run. |
| What does a Document Intelligence custom classifier do? | Identifies a document's type (class) so the correct extraction model can be invoked, useful for routing mixed-document batches. |
| Custom template vs custom neural model? | Template targets fixed/static layouts and trains fast with few samples; neural handles mixed/variable-layout documents but needs more data and training. |
| Read vs Layout model in Document Intelligence? | Read extracts printed/handwritten text (OCR) only; Layout adds tables, selection marks, paragraph roles, and structure, and can emit Markdown for RAG. |
| Which index field attributes increase storage, and which one does not? | `filterable`, `sortable`, and `facetable` add non-tokenized structures and increase storage; suggesters and edgeNgram tokenizers add more still. `retrievable` does not bloat the index — it only decides whether the field can be returned in results (and must be true for the key field). |
| When do you use `fieldMappings` versus `outputFieldMappings`? | `fieldMappings` maps a physical source field to an index field (name/type discrepancies, forked one-to-many paths, mapping functions). `outputFieldMappings` maps in-memory enrichment-tree nodes produced by skills to index fields. Enriched nodes never reach the index without an output field mapping or an index projection. |
| Define hybrid search in Azure AI Search. | A single request that runs full-text (BM25) and vector queries in parallel against one index and merges the results with Reciprocal Rank Fusion (RRF). |
| What do `projectionMode`, `parentKeyFieldName`, and `sourceContext` do in an index projection? | `projectionMode` is `skipIndexingParentDocuments` (recommended) or `includeIndexingParentDocuments` (default, adds parent rows with null chunk fields). `parentKeyFieldName` names the filterable field linking a chunk to its parent — never create a field mapping for it. `sourceContext` sets the granularity, for example `/document/pages/*`. |
| What do `maxFailedItems` and `maxFailedItemsPerBatch` control, and what are their defaults? | They set how many document-level errors an indexer tolerates before the run is marked Failed — overall and per batch. Both default to 0, so one bad document fails the entire run. Raise them to let a run report Success with per-item errors listed in execution history. |
| What are the limits and format of an Azure AI Search indexer `schedule`? | `interval` is an XSD dayTimeDuration (P(nD)T(nH)(nM)) with a minimum of 5 minutes and a maximum of 1,440 minutes (24 hours); `startTime` is optional UTC. Omit `schedule` and the indexer runs on demand only. If a run is still going when the next occurrence is due, that occurrence is skipped. |
| How do you read Get Indexer Status output? | The top-level `status` describes the indexer object — "running" means it is set up and available, not that a job is active. Each entry in `lastResult`/`executionHistory` has its own status (running, success, transientFailure, persistentFailure, reset) plus `itemsProcessed`, `itemsFailed`, errors, and warnings. History keeps the 50 most recent runs. |
| What components are required for integrated vectorization during indexing? | An indexer, a supported data source, and a skillset with a Text Split skill (chunking) and an Azure OpenAI Embedding skill (vectorization). |
| What is a knowledge store? | An optional output of a skillset that projects enriched documents as tables or blobs/objects in Azure Storage for downstream apps like Power BI and knowledge mining. |
| Which indexer mapping functions solve document key problems? | `base64Encode` makes a key URL-safe (set `useHttpServerUtilityUrlTokenEncode` true for keys); `base64Decode` reverses it with matching options; `fixedLengthEncode` shortens keys that exceed the 1,024-character limit; `urlEncode`/`urlDecode` convert only URL-unsafe characters. Mapping functions are not supported for index projections. |
| What is the queryFields add-on in Document Intelligence? | An add-on that lets you request extra fields beyond a model's default schema at analysis time, without training a custom model. |
| What algorithm fuses the keyword and vector result sets in a hybrid query? | Reciprocal Rank Fusion (RRF). |
| How do you invoke the semantic ranker in a query, and what does it do? | Set queryType=semantic with a semantic configuration; it applies machine reading comprehension to re-rank the top results for better relevance. |
| How does a skill's `context` change its execution? | Context is the scope of the operation. `/document` runs the skill once per document; adding `/*` to a collection node (for example `/document/pages/*`) invokes it once per item. Outputs are always added to the enrichment tree as children of the context node. |
| Name the main Split skill parameters and their effect. | `textSplitMode` (pages | sentences), `maximumPageLength` (300–50,000 characters, default 5,000), `pageOverlapLength` (characters/tokens copied from the previous chunk), `maximumPagesToTake` (0 = all), and `unit` (characters, default, or azureOpenAITokens). Output `textItems` is usually renamed with `targetName`. |
| Which two vector ranking algorithms does Azure AI Search support? | HNSW (Hierarchical Navigable Small World, approximate) and exhaustive KNN (eKNN). |
| What does vector search in Azure AI Search match on, and why does it find conceptually similar content? | It matches on numeric embeddings using a nearest-neighbors algorithm, so semantically similar items (e.g. 'dog' vs 'canine', or cross-language/cross-modal content) rank together even without keyword overlap. |
Use it as a check, not a first pass
If a line here is the first time you're meeting a fact, this sheet is doing the wrong job for you — go back to the domain notes and the quizzes. It works best as a rapid confirmation that recall is already automatic.