All certifications / Azure AI Engineer / Cheat sheet

Azure AI Engineer AI-102 cheat sheet

Every exam tip and key term from the free Azure AI Engineer lessons, by domain. Use your browser's Print to save it as a PDF.

Domain 1: Plan and manage an Azure AI solution (24%)

Exam tips

Match the output the scenario asks for: named fields from forms means Document Intelligence, plain text lines means Vision Read, harmful-content scores means Content Safety, and generated prose means Azure OpenAI.
If the question stresses one endpoint and one bill for several prebuilt services, pick the multi-service resource. If it stresses a Free tier or separate keys per service, pick single-service resources. Shared governance with separate workspaces means a hub with projects.
Global means any region, Data Zone means within the EU or US zone, Standard means the resource's region, and Provisioned means reserved capacity with predictable latency. A 429 error points to quota or rate limits.
In a template question, look at kind to identify the service and sku.name for the tier. customSubDomainName is the property tied to Entra ID authentication.
Know the status codes: 401 credential problem, 403 permission or network block, 404 wrong path or deployment name, 429 rate limit, 202 asynchronous operation started.
Reader is a management-plane role and cannot call AI APIs. For keyless calls choose a managed identity plus Cognitive Services User (or OpenAI User for Azure OpenAI). Token auth needs a custom subdomain.
Rotation without downtime is always switch to the other key, then regenerate. Keeping a resource off the internet is a private endpoint plus disabling public network access.
Know the three required container parameters: ApiKey, Billing and Eula. No internet at all means disconnected containers, which need Microsoft approval and a commitment plan.
Metrics are automatic and good for alerts; logs need a diagnostic setting and go to Log Analytics for KQL. Every alert notifies through an action group.
Map the scenario to the principle: explaining AI use and limits is transparency, equal treatment of groups is fairness, accessibility is inclusiveness, human ownership of outcomes is accountability. Specific forbidden words mean a blocklist.
400 with content_filter means the prompt was blocked; finish_reason content_filter means the output was blocked. Jailbreak in the user message is a user prompt attack; instructions hidden in data are a document attack.

Key terms

Azure AI services: Microsoft's family of prebuilt AI APIs for vision, language, speech, translation, documents and content safety, called over REST or SDKs.
Azure OpenAI: Azure-hosted OpenAI models for chat, reasoning, embeddings and image generation, deployed and managed through Azure AI Foundry.
Document Intelligence: A service that extracts text, structure and named fields from documents using prebuilt or custom models.
Content Safety: A service that detects harmful text and images across hate, sexual, violence and self-harm categories.
Single-service resource: An Azure resource for one AI service, with its own endpoint, keys, pricing tier and bill.
Multi-service resource: An Azure AI services resource that exposes many AI services behind one endpoint and key pair with combined billing.
Hub: A shared Azure AI Foundry container, built on Azure Machine Learning, that holds connections, security and compute for several projects.
Project: A Foundry workspace where you deploy models, build agents, evaluate and manage the assets for one app or team.
Deployment: A named instance of a model version inside an Azure OpenAI or Foundry resource, called by name from code.
Tokens per minute (TPM): The rate limit and quota unit for pay-as-you-go deployments.
Provisioned throughput unit (PTU): A unit of reserved model processing capacity for provisioned deployments.
Data zone: A group of regions, such as the EU or US, within which Data Zone deployments process requests.
kind: The template property that sets which AI service a Microsoft.CognitiveServices/accounts resource provides.
Bicep: A domain-specific language for Azure infrastructure as code that compiles to ARM JSON templates.
customSubDomainName: The property that gives a resource a unique endpoint name, required for Entra ID authentication and private endpoints.
disableLocalAuth: A property that turns off key-based authentication so only Microsoft Entra ID tokens work.
Endpoint: The base URL of an Azure AI resource that all API calls are sent to.
Ocp-Apim-Subscription-Key: The HTTP header that carries a resource key for most Azure AI services.
DefaultAzureCredential: An Azure Identity class that finds a Microsoft Entra ID credential automatically from the environment, managed identity or developer sign-in.
Operation-Location: The header returned by asynchronous operations that points to the URL you poll for the result.
Managed identity: An Entra ID identity for an Azure resource whose credentials Azure creates and rotates automatically.
Cognitive Services User: An RBAC role that lets an identity call Azure AI services data-plane APIs.
Cognitive Services OpenAI User: An RBAC role that allows inference calls to Azure OpenAI deployments without management rights.
Custom subdomain: A unique resource-specific endpoint name required for Entra ID token authentication.
Azure Key Vault: A managed service for storing and auditing secrets, keys and certificates.
Key rotation: Replacing keys regularly, using the two-key pattern so clients never lose access.
Private endpoint: A network interface with a private IP in your virtual network that connects privately to an Azure resource.
Customer-managed key: An encryption key you own in Key Vault used to encrypt a service's data at rest instead of a Microsoft-managed key.
Connected container: A container that processes data locally but reports usage to Azure for billing.
Disconnected container: A container approved for fully offline use under a commitment plan, with no online usage reporting.
Billing setting: The container startup value that holds the endpoint of the Azure resource charged for usage.
Eula=accept: The required startup argument confirming acceptance of the container's license terms.
Diagnostic setting: A resource configuration that routes logs and metrics to Log Analytics, storage or an event hub.
Action group: A reusable set of notifications and automated actions triggered by alerts.
Commitment tier: A pricing plan with a fixed monthly fee for a set volume of usage at a discounted rate.
Budget: A Cost Management threshold on actual or forecasted spend that sends alerts.
Harm categories: The four Content Safety classifications: hate, sexual, violence and self-harm.
Severity level: A score showing how harmful content is within a category, which your app compares to a threshold.
Blocklist: A custom list of terms that Content Safety flags in addition to its harm classifiers.
Transparency: The responsible AI principle that people should know they are using AI and understand its capabilities and limits.
Content filter: The configurable input and output moderation applied to every Azure OpenAI deployment.
Prompt shields: A Content Safety feature that detects user prompt attacks (jailbreaks) and document attacks (indirect prompt injection).
Groundedness detection: A check that flags model output not supported by the provided source material.
Indirect prompt injection: Malicious instructions hidden in content the model reads, such as documents or tool results.

Domain 2: Implement generative AI solutions (18%)

Exam tips

Vectors for search means an embedding model; new pictures means an image model; hard multistep logic suggests a reasoning model; everything conversational is a chat model. Code calls the deployment name, not the model name.
Rules and persona go in the system message. Memory in a chat app comes from resending prior messages, not from the deployment. finish_reason length means raise max tokens.
Consistency means lower temperature. Cut-off answers with finish_reason length mean raise max tokens. Repetition means frequency or presence penalty. Change temperature or top_p, not both.
When the scenario is about getting an exact format or style, the answer is usually few-shot examples or structured output. Instructions and rules belong in the system message; content should be clearly delimited.
Up-to-date facts, private documents and citations point to RAG, not fine-tuning. If answers are wrong because the right passage is missing, fix retrieval (hybrid search, semantic ranking, better chunking) before changing the model.
The vector field's dimensions must match the embedding model, and queries must be embedded with the same model as the documents. Chunk overlap prevents losing context at boundaries.
Comparing prompt versions in one node means variants; storing endpoints and keys means connections; a conversational flow with history is a chat flow; publishing it as an API means a managed online endpoint.
Answers not supported by retrieved context point to groundedness; off-topic answers point to relevance; poor readability points to fluency or coherence. Overlap metrics such as F1 and BLEU need ground truth.
Facts that change and need citations: RAG. Consistent style or format, or shorter prompts on a narrow task: fine-tuning. Try prompt engineering first in every case.
Creating pictures means an image generation model; asking questions about a picture means sending it as an image part to a vision-enabled chat model. For high-volume structured extraction, prefer Document Intelligence or Vision.
429 means rate limits: back off and retry, add quota or deployments, or use provisioned throughput. Finding which step caused a slow or bad answer means tracing with Application Insights.

Key terms

Model catalog: The Azure AI Foundry list of models from Microsoft, OpenAI and partners, with model cards and deployment options.
Reasoning model: A model that works through a problem internally before answering, improving accuracy on complex tasks at higher latency and cost.
Embedding model: A model that converts text into a numeric vector representing its meaning, used for search and similarity.
Context window: The maximum number of tokens a model can process across input and output in one request.
System message: The instruction message that sets the model's behavior, rules and format for the whole conversation.
Stateless API: An API that keeps no memory between calls, so the client must resend context each time.
finish_reason: A field explaining why generation stopped: stop, length, content_filter or tool_calls.
Streaming: Returning generated tokens incrementally as they are produced instead of in one final response.
Temperature: A parameter from 0 to 2 that controls randomness; lower values give more deterministic output.
top_p: Nucleus sampling: only tokens within the top cumulative probability p are considered.
Stop sequence: A string that ends generation when the model produces it.
Frequency penalty: A parameter that reduces the likelihood of tokens in proportion to how often they have already appeared.
Few-shot prompting: Including example inputs and desired outputs in the prompt to show the model the task and format.
Zero-shot prompting: Asking the model to perform a task with instructions only and no examples.
Chain-of-thought: Prompting the model to reason step by step before giving its final answer.
Delimiter: Markers such as triple quotes or tags that separate instructions from the content being processed.
RAG: Retrieval augmented generation: retrieving relevant data at query time and adding it to the prompt to ground the answer.
Grounding: Supplying the model with source content it should base its answer on.
Ingestion: The offline process of extracting, chunking, embedding and indexing documents for retrieval.
Citation: A reference in the answer to the source document or chunk that supported it.
Embedding: A numeric vector representation of text in which similar meanings are close together.
Cosine similarity: A measure of how similar two vectors are based on the angle between them.
Chunking: Splitting documents into smaller passages, often with overlap, before embedding and indexing.
HNSW: An approximate nearest neighbor algorithm that finds similar vectors quickly in large indexes.
Flow: A directed graph of tool nodes that defines a generative AI pipeline in prompt flow.
Variant: An alternative version of an LLM node's prompt or parameters, run side by side for comparison.
Connection: A stored, reusable configuration of an external resource's endpoint and credentials used by flow nodes.
Evaluation flow: A flow that scores another flow's outputs against ground truth or quality criteria.
Groundedness: An evaluation metric scoring whether the answer's claims are supported by the provided context.
Relevance: An evaluation metric scoring whether the answer addresses the user's question.
Ground truth: The expected correct answer for a test input, used by similarity and overlap metrics.
Red teaming: Deliberately attacking a system with adversarial inputs to find safety weaknesses before real users do.
Fine-tuning: Further training a base model on your own examples to produce a customized model version.
JSONL: A file format with one JSON object per line, used for fine-tuning training data.
Epoch: One full pass of training over the whole training dataset.
Validation file: Held-out examples used during fine-tuning to measure how well the model generalizes.
Image generation model: A model such as DALL-E 3 that creates new images from a text prompt.
Multimodal model: A model that accepts more than one type of input, such as text and images, in the same request.
Revised prompt: The expanded prompt DALL-E 3 actually used, returned alongside the generated image.
Detail setting: An option that controls how much of an input image the model processes, trading cost for accuracy.
Prompt tokens: Input tokens in a request, including system message, history and retrieved context.
Exponential backoff: Retrying failed requests with increasing wait times to avoid overwhelming a service.
Tracing: Recording each step of a request with inputs, outputs, timing and token use for debugging.
AI gateway: A layer such as Azure API Management in front of model deployments that balances load, enforces limits and logs usage.

Domain 3: Implement an agentic solution (8%)

Exam tips

Multi-step tasks that call APIs or take actions suggest an agent. Single-shot text transformations (rewrite, translate, summarize one document) suggest a plain chat completion.
Conversation history lives in the thread, behavior in the agent, and processing in the run. requires_action means your code must run the function and submit tool outputs.
Calculations or charts from uploaded data: code interpreter. Answers from uploaded documents: file search. Existing enterprise index: Azure AI Search tool. Existing REST API with a spec: OpenAPI. Your own code in your app: function calling.
The model only proposes calls; your code executes them and must send results back with the matching tool call ID. Descriptions drive which function the model chooses.
Exposing your own methods to the model in Semantic Kernel means a plugin with kernel functions. AutoGen is associated with multi-agent conversations. Agent Service is the managed, hosted option.
A main agent delegating to specialists maps to connected agents or handoff. Agents reviewing each other's work is group chat or maker-checker. Fixed step order is sequential orchestration.
Risky or irreversible actions need human approval enforced in code. Limit damage with least-privilege tool permissions. Finding out which tool call went wrong needs tracing.
Apps call hosted agents through the project endpoint with Entra ID, ideally a managed identity. Framework-based agents are deployed like any app (Container Apps, App Service, Functions).

Key terms

AI agent: Software that uses a model to plan and take actions through tools in a loop to achieve a goal.
Tool: A capability an agent can invoke, such as a function, API, search index or code interpreter.
Instructions: The agent's system-level guidance defining its role, goals, rules and behavior.
Thread: Stored conversation state, including messages and tool results, that an agent works from across turns.
Agent: In Agent Service, a definition combining a model deployment, instructions and tools.
Run: One activation of an agent on a thread that processes messages and tool calls and appends replies.
requires_action: A run status meaning the app must execute requested function calls and submit their outputs.
Run step: An individual action within a run, such as a tool call or message creation, used for inspection and debugging.
File search: A built-in agent tool that indexes uploaded files in a vector store and retrieves passages with citations.
Code interpreter: A built-in agent tool that writes and runs Python in a sandbox for calculations, data processing and charts.
OpenAPI tool: A tool that lets an agent call a REST API described by an OpenAPI specification.
Vector store: The managed store of chunked, embedded files that file search queries.
Function calling: A model feature that returns a structured request to call a described function with arguments, which the app executes.
JSON schema: A standard for describing the structure, types and required fields of JSON data, used for function parameters.
tool_call_id: The identifier linking a tool result message to the model's specific tool call request.
tool_choice: A parameter that lets the model choose tools automatically, forces a specific one or disables them.
Semantic Kernel: Microsoft's open-source SDK for integrating AI models, plugins and agents into C#, Python and Java apps.
Kernel: The Semantic Kernel container holding AI service connections and plugins.
Plugin: A group of functions, described for the model, that Semantic Kernel exposes as callable tools.
AutoGen: A Microsoft Research open-source framework for building conversations between multiple agents.
Orchestration: Coordinating multiple agents so their work combines to complete a task.
Handoff: A pattern where one agent transfers control of a conversation to a more suitable agent.
Connected agents: An Agent Service feature that lets a main agent call specialist agents as tools.
Maker-checker: A pattern where one agent produces work and another reviews it, iterating until accepted.
Least privilege: Granting an agent and its tools only the minimum permissions needed for their task.
Human-in-the-loop: Requiring a person to review or approve an AI action before it takes effect.
Prompt injection: An attack that inserts instructions into model input to override the intended behavior.
OpenTelemetry: An open standard for traces, metrics and logs, used to trace agent runs into Application Insights.
Project endpoint: The Foundry project URL that apps use with the SDK to call Agent Service.
Thread mapping: Storing each user's thread ID so their conversation can continue across requests.
Quality gate: An automated evaluation that must pass before a change is promoted to the next environment.
Standard setup: An Agent Service configuration that uses your own storage, search and networking resources for agent data.

Domain 4: Implement computer vision solutions (13%)

Exam tips

One sentence for the whole image is caption; sentences with boxes for regions is dense captions; single words are tags; where things are is objects; detecting people is not identifying them.
Text in photos and scenes points to Vision Read; long documents, tables and form fields point to Document Intelligence. Read returns lines and words with polygons and word confidence.
Exactly one label per image: multiclass. Several labels per image: multilabel. Location or count: object detection. Offline or edge: compact domain.
Precision is about how many predictions were right; recall is about how many real cases were found. Raising the threshold trades recall for precision. mAP is the headline metric for object detection.
Retrained but no change in the app: the new iteration was not published under the name the app calls. Offline use: export, which requires a compact domain.
Detection and quality attributes are available; identification and verification need Limited Access approval; emotion, age and gender inference were retired.
Searchable transcripts, on-screen text, topics and scenes from recorded videos point to Video Indexer. Recognizing specific faces in video still requires Limited Access approval.
General concepts, no training: Image Analysis. Your own categories at scale or offline: Custom Vision. Open-ended questions or reasoning about an image: a multimodal model. Forms: Document Intelligence.

Key terms

Caption: An Image Analysis feature that returns one sentence describing the whole image.
Dense captions: An Image Analysis feature that returns captions with bounding boxes for multiple regions of an image.
Tags: Single-word labels for content in an image, each with a confidence score.
Smart crops: Suggested crop regions that preserve the most important part of an image for given aspect ratios.
OCR: Optical character recognition: extracting machine-readable text from images of text.
Read feature: The Azure AI Vision capability that extracts printed and handwritten text with positions and confidence.
Bounding polygon: The corner points outlining where a line or word appears in the image, including rotated text.
Line and word: The levels at which Read returns text, each with position; words also include confidence scores.
Multiclass classification: A classification type where each image gets exactly one tag.
Multilabel classification: A classification type where each image can have any number of tags.
Object detection: A model type that returns a tag and bounding box for each instance of an object in an image.
Compact domain: A Custom Vision base model that produces smaller, exportable models for offline and edge use.
Iteration: A versioned model produced by one Custom Vision training run.
Precision: The fraction of the model's positive predictions for a tag that were correct.
Recall: The fraction of actual instances of a tag that the model found.
mAP: Mean average precision: the average of per-tag average precision, summarizing detection quality.
Publish name: The name under which an iteration is published to a prediction resource and called by apps.
Prediction resource: The Custom Vision resource that hosts published iterations and serves prediction requests.
Prediction-Key: The HTTP header carrying the prediction resource's key when calling a published model.
Export: Downloading a compact-domain model in a format such as ONNX, TensorFlow, Core ML or a Docker container for offline use.
Face detection: Finding faces in an image and returning their locations, with optional landmarks and attributes.
Face verification: A one-to-one check of whether two faces belong to the same person; a Limited Access feature.
Face identification: A one-to-many search matching a face against enrolled people; a Limited Access feature.
Limited Access: Microsoft's approval process required before using sensitive AI capabilities such as face recognition.
Video Indexer: An Azure AI service that extracts time-coded audio and visual insights from video.
Insights: The JSON output of indexing, such as transcripts, OCR, labels, keywords, faces and scenes, each with timestamps.
Widgets: Embeddable player and insights components for showing indexed video in your own web app.
Keyframe: A representative frame chosen from a shot, used for thumbnails and visual analysis.
Prebuilt model: A model trained by Microsoft that works without your training data.
Custom model: A model trained on your own labeled data to recognize your specific categories.
Multimodal model: A model that accepts images with text and can reason and answer in natural language.
Edge deployment: Running a model on a local device rather than calling a cloud endpoint.

Domain 5: Implement natural language processing solutions (18%)

Exam tips

Overall positive or negative is sentiment; which aspect is liked or disliked is opinion mining. Categorizing entities is NER; resolving which real thing an entity refers to is entity linking.
Mask personal data in text: PII detection with redactedText. Restrict which types are masked with the categories filter. Healthcare data: domain phi.
Several target languages in one call: multiple to parameters. 401 with a regional key: add the Ocp-Apim-Subscription-Region header. Whole files with formatting: document translation. Domain terminology: Custom Translator with a category ID.
One short command: recognize once. Long live audio: continuous recognition with events. Thousands of stored files: batch transcription. A Canceled result usually means credential, region or connection problems.
Speed and pitch: prosody. Pauses: break. Dates, numbers and spelling: say-as. Exact pronunciation: phoneme or a lexicon. Emotion or style: mstts:express-as.
Live speech into another language: SpeechTranslationConfig with add_target_language and a TranslationRecognizer. A wake word: keyword recognition with a custom keyword model. Understanding commands: CLU intents and entities.
Out-of-scope input landing in a real intent: add examples to None. A fixed set of values with synonyms: list entity. Patterns like codes: regex entity. Values that vary in context: learned entity.
Clarifying follow-ups: multi-turn prompts. Equivalent words across the whole project: synonyms. Wrong answers to unrelated questions: raise the confidence threshold and set a default answer. It needs an Azure AI Search resource.
One category per document: single-label. Any number: multi-label. Your own entity types from long documents: custom NER. Training data lives in a Blob Storage container connected to the Language resource.
Vocabulary problems: plain text data or a phrase list. Noise or accent problems: audio plus human-labeled transcripts. Custom neural voice needs Limited Access approval and the talent's recorded consent.

Key terms

Language detection: Identifying the language of text, returned as a name, ISO code and confidence score.
Named entity recognition: Finding entities in text and classifying them into categories such as Person, Location and Organization.
Entity linking: Disambiguating an entity by linking it to a specific knowledge base entry.
Opinion mining: Aspect-based sentiment that connects sentiment to specific targets mentioned in text.
PII: Personally identifiable information: data that can identify a person, such as names, phone numbers or ID numbers.
Redaction: Masking or replacing sensitive values in text so they are not exposed.
redactedText: The PII detection output field containing the input text with detected entities masked.
PHI: Protected health information, detected by setting the PII domain to phi.
Text translation: Real-time translation of strings through the Translator translate operation.
Document translation: Asynchronous translation of whole files in Blob Storage that preserves formatting.
Custom Translator: A service for training translation models on your own parallel documents, used by category ID.
Transliteration: Converting text from one script to another without translating its meaning.
SpeechConfig: The Speech SDK object holding credentials, region or endpoint, and settings such as recognition language.
recognize_once_async: A Speech SDK method that recognizes a single utterance, ending at a pause.
Continuous recognition: Event-driven recognition of long audio, with interim and final results, until stopped.
Batch transcription: An asynchronous REST API for transcribing large sets of stored audio files.
Neural voice: A prebuilt synthetic voice generated by deep neural networks for natural-sounding speech.
SSML: Speech Synthesis Markup Language, XML that controls voice, pronunciation, pauses, rate and style.
prosody: The SSML element that adjusts speaking rate, pitch and volume.
say-as: The SSML element that controls how values like dates, numbers and letters are read.
SpeechTranslationConfig: The Speech SDK configuration for translating speech, with source language and target languages.
TranslationRecognizer: The Speech SDK recognizer that returns recognized text and its translations.
Intent recognition: Determining the purpose of an utterance and extracting its entities.
Keyword recognition: On-device detection of a wake word using a custom keyword model before full recognition starts.
Intent: The goal or action a user expresses in an utterance, such as BookFlight.
Entity: A piece of information in an utterance the app needs, such as a date or destination.
Utterance: An example phrase a user might say, used to train and test the model.
None intent: The built-in intent that captures utterances outside the app's scope.
Project: A custom question answering knowledge base of question and answer pairs built from sources.
Follow-up prompt: A link from an answer to related pairs, enabling multi-turn conversations.
Synonyms: Project-level word alternatives treated as equivalent when matching questions.
Confidence threshold: The minimum score an answer needs to be returned; below it the default answer is used.
Single-label classification: Custom text classification where each document receives exactly one class.
Multi-label classification: Custom text classification where each document can receive zero or more classes.
Custom NER: A trained model that extracts your own entity types from unstructured documents.
F1 score: The harmonic mean of precision and recall, a single measure of model accuracy.
Custom speech: A speech to text model adapted with your text and audio data for better accuracy in your domain.
Word error rate (WER): The share of words substituted, deleted or inserted compared with a reference transcript.
Custom neural voice: A Limited Access feature that trains a synthetic voice resembling a specific consenting voice talent.
Phrase list: A runtime list of words that boosts their recognition without training a custom model.

Domain 6: Implement knowledge mining and information extraction solutions (19%)

Exam tips

Slow queries or availability: add replicas. Bigger index or slow indexing: add partitions. Pull from Azure data on a schedule: indexer. Plan the tier up front: downgrades are not possible and in-place upgrades are limited to eligible services.
Map the query feature to the attribute: $filter needs filterable, $orderby needs sortable, facets need facetable, full-text needs searchable, returned in results needs retrievable. Autocomplete needs a suggester.
Scanned text needs imageAction generateNormalizedImages plus the OCR skill. Enriching more than a handful of documents needs an attached Azure AI services resource. Enriched values reach the index through output field mappings.
Your own model or lookup during indexing means a custom Web API skill. The endpoint must accept and return a values array with matching recordIds and a data object.
Fuzzy, proximity, boosting, regex and fielded queries need queryType full. Filters are OData with word operators (eq, ge, and), not symbols like == or &&.
Exact codes plus meaning: hybrid search. Better ordering of the top results and extracted answers: semantic ranking with a semantic configuration and queryType semantic. Queries must use the same embedding model as the documents.
Analytics in Power BI: table projections. Hierarchical JSON for other systems: object projections. Extracted images: file projections. Related tables that must join: same projection group.
Text only: read. Tables, checkboxes and structure without named fields: layout. Named fields from common documents: the matching prebuilt model. 202 plus Operation-Location means poll for the result.
Fixed layout: template. Varying layouts from many sources: neural. Mixed document types in one stream: custom classifier. One model ID for several custom models: composed model.
One service with a field schema across documents, images, audio and video points to Content Understanding. The analyzer holds the schema; extract, generate and classify are the field methods.

Key terms

Index: The searchable collection of documents and its schema of fields in Azure AI Search.
Indexer: A crawler that reads a data source, cracks documents, maps fields, runs skillsets and loads the index.
Replica: A copy of the index that serves queries, added for throughput and availability.
Partition: A unit of storage and indexing capacity, added for larger indexes and faster indexing.
Key field: The single Edm.String field that uniquely identifies each document in an index.
Facetable: A field attribute that enables counts of documents per value for faceted navigation.
Analyzer: A component that tokenizes and normalizes text for full-text search, such as a language analyzer.
Suggester: An index definition that enables autocomplete and search-as-you-type suggestions on chosen fields.
Skillset: A collection of AI enrichment skills that an indexer runs during indexing.
Document cracking: The indexer step that opens files and extracts text, metadata and images.
Enrichment tree: The in-memory structure holding a document's original and enriched data as skills run.
Output field mapping: A mapping that copies enriched values from the enrichment tree into index fields.
Custom skill: A skill that calls your own code, such as an Azure Function, during AI enrichment.
WebApiSkill: The skill type that calls an external HTTPS endpoint with a defined JSON interface.
recordId: The identifier that links each input record to its output in the custom skill interface.
values array: The top-level array of records sent to and returned from a custom skill.
queryType: The query parser: simple (default), full (Lucene) or semantic.
$filter: An OData expression that restricts results by filterable field values without affecting scoring.
Fuzzy search: A full Lucene query that matches terms with small spelling differences, written with ~.
Facets: Counts of matching documents per field value or range, used for faceted navigation.
Vector search: Finding documents whose embedding vectors are closest to the query's vector.
Hybrid search: Combining keyword and vector queries in one request, merged with Reciprocal Rank Fusion.
Reciprocal Rank Fusion: A method that merges ranked lists by scoring documents on their rank positions in each list.
Semantic ranker: A second-stage model that re-ranks top results by meaning and can return captions and answers.
Knowledge store: Azure Storage output of a skillset's enriched data, saved for uses beyond search.
Table projection: A projection that writes enriched data as rows in Azure Table Storage.
Object projection: A projection that writes enriched data as JSON documents in Blob Storage.
Shaper skill: A utility skill that builds a custom data shape from the enrichment tree for projection.
Read model: The Document Intelligence model that extracts text lines and words from documents.
Layout model: The model that extracts text plus structure: paragraphs, tables, selection marks and figures.
Prebuilt model: A ready-to-use model that extracts named fields from a common document type such as invoices or receipts.
Confidence score: A value from 0 to 1 indicating how certain the model is about an extracted field.
Custom template model: A custom extraction model for documents with a consistent, fixed layout.
Custom neural model: A deep learning custom extraction model that handles documents with varying layouts.
Custom classifier: A model that identifies document types, and splits combined files, before extraction.
Composed model: Several custom models grouped under one model ID that routes each document to the best match.
Content Understanding: An Azure AI service that extracts structured, schema-defined output from documents, images, audio and video.
Field schema: The list of named, typed and described fields an analyzer should return.
Generate field: A schema field whose value the model produces from the content, such as a summary.

Study Azure AI Engineer for free
Lessons, quizzes, exam simulations and hands-on labs.

Open the Azure AI Engineer study plan