Microsoft Azure AI Cloud Developer Associate — knowledge map

The 111 core concepts of Microsoft Azure AI Cloud Developer Associate and how they connect. Click a node in the map above to explore related terms and prerequisites; the list below indexes every concept with its definition and links to its prerequisites and related concepts.

Concepts (111)

• Generative AI

  AI that creates new content (text, images, code) from learned data; distinct from classification/regression.

• Azure Machine Learning

  Supports the ML workflow (prepare→train→evaluate→deploy→predict); includes Automated ML (auto-best) and Designer (no-code).

  Prerequisites: Machine learning (ML)

• Azure ML deployment (online/batch endpoints, traffic split)

  Inference is served via endpoints. A managed online endpoint = real-time low-latency (Azure manages infra, autoscale); Kubernetes online = existing AKS; batch endpoint = async scoring of large data on a compute cluster. Place multiple deployments under one endpoint and use traffic split for blue/green (gradual rollout, instant rollback). A deployment needs a scoring script (init/run), an environment, and the right instance type/count (MLflow models may need no inference code). Access via managed identity + key/token auth, with private endpoints if needed.

  Prerequisites: Azure Machine Learning、Batch endpoint、Compute cluster、Managed online endpoint

• Azure ML jobs and AutoML/sweep (MLflow, hyperparameter tuning)

  Training runs as jobs: command (single script), sweep (hyperparameter search), and pipeline (a DAG of components). Define in YAML (CLI v2)/SDK v2. Tracking uses natively integrated MLflow (log_metric/log_param/log_artifact, autolog) to compare experiments. AutoML auto-searches algorithms/preprocessing from a task and picks via a leaderboard. A sweep tunes hyperparameters with a search space + sampling (grid/random/Bayesian) + early termination (Bandit/median stopping), controlling parallelism/cost via max_concurrent_trials.

  Prerequisites: Azure Machine Learning、Command job、Component (reusable step)、Early termination policy (Bandit/median)

• Azure OpenAI Service

  Securely access foundation models (GPT) via Azure to build generative AI into your apps (enterprise controls).

  Prerequisites: Foundation model、Generative AI

• Azure ML compute and data (workspace, cluster, data assets)

  The Azure ML workspace centrally manages ML artifacts (data, models, experiments, endpoints) with associated resources (storage/Key Vault/Application Insights/Container Registry). Compute: a compute instance for dev (personal; stop to save cost) and a compute cluster to scale training (autoscale, 0 nodes when idle, low-priority/Spot VMs). Data: a datastore (connection info to Blob/ADLS) and data assets (reusable, versioned references: uri_file/uri_folder/mltable). Trained models are versioned in the model registry. Environments (conda/Docker) pin dependencies for reproducibility. Interact via Studio/Python SDK v2/CLI v2.

  Prerequisites: Azure Machine Learning、Compute cluster、Compute instance、Data assets (uri_file/uri_folder/mltable)

• Embeddings

  Numeric vector representations of text (etc.) enabling comparison of semantic closeness; used in search and RAG.

• Token

  The unit of text a model processes (roughly part of a word); basis for cost and context length.

• Semantic ranker and vector search

  A semantic ranker that reorders results by meaning, and vector search that finds similarity via embeddings. Combining both with keywords (hybrid search) improves RAG quality.

  Prerequisites: Embeddings

• Event-driven AI pipelines

  A design where Event Grid/Service Bus/Event Hubs trigger asynchronous AI processing (ingest → embed → index → infer) via events and messages.

  Prerequisites: Embeddings、Azure Event Grid、Azure Event Hubs、Azure Service Bus

• Azure AI Language

  Text understanding: sentiment, entity recognition, key phrases, language detection, summarization, CLU.

  Prerequisites: Conversational Language Understanding (CLU)、Key phrase extraction and language detection、Entity recognition (NER) and PII detection、Sentiment analysis and opinion mining

• Azure Container Apps

  A serverless platform to run containers/microservices without managing Kubernetes; KEDA-based event-driven scaling (scale to zero) and Dapr integration.

  Prerequisites: ML data preparation (encoding/scaling/imbalance)

  Related: ACR (Azure Container Registry)

• Sweep job (hyperparameter tuning)

  A job that runs many trials with a search space, sampling, primary metric, and early termination to find the best hyperparameter configuration.

  Prerequisites: Search space

  Related: Primary metric

• ML data preparation (encoding/scaling/imbalance)

  Shaping raw data for learning. Encoding turns categories numeric (one-hot/ordinal/target); scaling = standardize (z-score)/normalize (Min-Max); imbalance = SMOTE oversampling/class weights (evaluate with F1/PR-AUC). Fit transforms only on post-split training data (prevent leakage).

• ACR (Azure Container Registry)

  A private registry to store/manage container images; ACR Tasks builds in the cloud, and authentication uses a managed identity.

  Prerequisites: Managed identity

  Related: Azure Container Apps

• Vector data management for AI

  Stores embedding vectors for similarity search. On Azure, choose among Cosmos DB (NoSQL vector search), PostgreSQL pgvector, and Azure AI Search vector indexes by use case.

  Prerequisites: Semantic ranker and vector search、Azure AI Search (knowledge mining)、Azure Cosmos DB、Embeddings

• Azure AI Search (knowledge mining)

  Indexes large document sets for cross-document search (knowledge mining); also a retrieval backend for RAG.

• Azure Cosmos DB

  A globally distributed, low-latency, auto-scaling managed NoSQL DB supporting multiple APIs (NoSQL/MongoDB/Cassandra/Gremlin/Table).

  Prerequisites: ML data preparation (encoding/scaling/imbalance)

• Component (reusable step)

  A reusable unit of processing bundling inputs, outputs, code, and environment. Composed as pipeline steps and versioned/shared.

• Foundation model

  A large pretrained model usable across many tasks (e.g., GPT); multimodal handles multiple data types.

• Azure Key Vault

  Securely stores and manages keys, secrets, and certificates so they need not be embedded in apps.

• Managed identity

  An identity letting apps access Azure resources securely without managing secrets; a service principal auto-managed by Azure.

• Document Intelligence custom/composed models

  Custom models trained on your own forms (template or neural) and a composed model that bundles multiple custom models and routes automatically.

  Prerequisites: Azure AI Document Intelligence

• Speech-to-text (STT) and custom speech

  Speech recognition that converts spoken words to text. Custom Speech tailors recognition to jargon or accents to improve accuracy (distinct from custom neural voice on the text-to-speech side).

  Prerequisites: Document Intelligence custom/composed models、Text-to-speech (TTS) and SSML

• Text-to-speech (TTS) and SSML

  Text-to-speech converting text to natural audio. SSML (Speech Synthesis Markup Language) controls pronunciation, prosody, and pauses, and custom neural voice creates a bespoke voice.

  Prerequisites: Document Intelligence custom/composed models

• RAG optimization

  RAG quality depends on retrieval. Tune similarity threshold, chunk size, and retrieval strategy (top-k / re-ranking); for domain specificity, select/fine-tune the embedding model. Complement with hybrid search combining semantic (vector) and keyword retrieval, and evaluate/improve quantitatively with relevance metrics and A/B testing.

  Prerequisites: Semantic ranker and vector search、Generative AI quality metrics and evaluation、Embeddings

• Azure Kubernetes Service (AKS)

  Managed Kubernetes for production container orchestration (placement, self-healing, scaling).

  Prerequisites: ML data preparation (encoding/scaling/imbalance)

• Azure API Management (APIM)

  An API gateway exposing multiple backend APIs through one front door. Policies (XML; inbound/backend/outbound/on-error) apply auth (validate-jwt), rate limiting, transformation, caching, and CORS. Offer to consumers via products and subscriptions (keys), document in the developer portal, authenticate to backends with a managed identity, and store secrets in named values (Key Vault).

  Prerequisites: Azure Key Vault、Managed identity

• Azure AI Speech

  Speech ⇄ text (speech-to-text, text-to-speech) and speech translation.

  Prerequisites: Speech translation and speaker recognition、Speech-to-text (STT) and custom speech、Text-to-speech (TTS) and SSML

• Azure AI Vision

  Prebuilt general computer vision: tags, captions, object detection, OCR (Read).

  Related: OCR (Read)

• Compute cluster

  Autoscaling compute that scales training or batch inference. It can shrink to 0 nodes when idle and use low-priority/Spot VMs to cut cost.

• Model catalog

  A catalog in Azure AI Foundry / Azure ML to discover, compare, and deploy many foundation models (OpenAI, Hugging Face, Meta, etc.).

  Prerequisites: Azure Machine Learning、Foundation model

• Prompt flow

  A development tool that links LLMs, prompts, Python, and tools as nodes to build, evaluate, and deploy generative-AI app flows. Chaining logic is defined with the SDK.

  Prerequisites: Generative AI

• Search space

  The definition of which hyperparameters to tune and their candidate ranges (discrete choice, continuous uniform/loguniform, etc.).

• Azure ML workspace

  The top-level resource that centrally manages ML artifacts (data, models, experiments, endpoints), with associated storage, Key Vault, Application Insights, and Container Registry.

  Prerequisites: Azure Machine Learning、Azure Key Vault

• Azure Event Grid

  Reactive pub/sub delivering discrete events (e.g., "blob created"); subscribe via event subscriptions and react in near real time.

• Azure Event Hubs

  High-throughput ingestion of large streaming/telemetry; processed in parallel via partitions and consumer groups.

• GenAIOps

  An operational practice extending MLOps concepts to generative AI (Azure AI Foundry, foundation models, prompts). Prompts, agent flows, and evaluation metrics (groundedness, harmfulness, etc.) are managed as code/config in Git, with CI/CD automating the evaluate-deploy-monitor-improve cycle.

  Prerequisites: Foundation model、Generative AI、MLOps (SageMaker Pipelines / Model Registry)

• Responsible AI

  Guidance for using AI safely, fairly, transparently; Microsoft’s six: fairness, reliability & safety, privacy & security, inclusiveness, transparency, accountability.

• Azure Service Bus

  An enterprise message broker supporting ordering, transactions, dedup, and dead-lettering; used for reliable command delivery.

• Content filters and blocklists

  Safety features of Azure OpenAI / Content Safety: filter by per-category severity thresholds and add banned terms via a custom blocklist.

  Prerequisites: Azure AI Content Safety、Azure OpenAI Service

• Prompt flow (Foundry)

  A development feature linking LLMs, prompts, Python, and tools as nodes to build, evaluate, and deploy generative-AI flows, used to implement RAG and agents.

  Prerequisites: Prompt flow、Generative AI

• Index and indexer

  The index that defines what is searchable in Azure AI Search and the indexer that pulls content from data sources to populate it. Queries support syntax, sorting, filtering, and wildcards.

  Prerequisites: Azure AI Search (knowledge mining)

• Summarization (extractive/abstractive)

  Summarizes long text or conversations. Extractive summarization selects key sentences; abstractive summarization rephrases into new sentences.

• Containerized/serverless compute for AI

  Execution platforms for inference and AI pipelines. Azure Container Apps (scalable, KEDA) and Azure Functions (event-driven) run model calls and pre/post-processing serverlessly.

  Prerequisites: Azure Functions、Azure Container Apps

• Generative AI quality metrics and evaluation

  Because generative AI output is non-deterministic, evaluation is central. AI quality metrics = groundedness (based on provided context = opposite of hallucination) / relevance (answers the question) / coherence (logic) / fluency (naturalness). Add risk/safety evaluation (harmful content/jailbreak), made into a quality gate via automated evaluation workflows with built-in + custom metrics.

  Prerequisites: Generative AI

• Prompt flow CI/CD

  Treats prompt flow as code, building pipelines with evaluation-gated automated testing, approval, and deployment to iterate generative-AI apps safely.

  Prerequisites: Prompt flow、Generative AI

• Azure Functions

  The flagship serverless (FaaS): event-driven function execution billed only for what runs.

• Azure AI Content Safety

  A content filter that detects/suppresses harmful or inappropriate content; used to operate generative AI safely.

  Prerequisites: Generative AI

• Azure AI Document Intelligence

  Extracts structured fields (amounts, dates) from invoices, receipts, and forms (document intelligence).

• Batch endpoint

  An endpoint that asynchronously scores large data on a compute cluster. Invoking it starts a batch scoring job.

  Prerequisites: Compute cluster

• Early termination policy (Bandit/median)

  A sweep setting that cancels unpromising trials to save cost. Options include Bandit (slack factor), median stopping, and truncation selection.

  Prerequisites: Sweep job (hyperparameter tuning)

• MLflow tracking

  Experiment tracking natively integrated in Azure ML. Logs training via log_metric/log_param/log_artifact and autolog to compare runs.

  Prerequisites: Azure Machine Learning

• Model registry

  A place to version, register, and reference trained models with stages and lineage, used to select what to deploy.

• Pipeline job

  A sequence connecting multiple components as a DAG, passing data between steps. It can be scheduled and monitored.

  Prerequisites: Component (reusable step)

• Responsible AI dashboard

  A set of Azure ML tools to assess models for responsible AI, offering error analysis, interpretability, fairness, and counterfactuals in one dashboard.

  Prerequisites: Azure Machine Learning、Responsible AI

• Sampling method (grid/random/Bayesian)

  How a sweep picks configurations from the search space: grid (exhaustive), random (broad and fast), and Bayesian (learns promising regions from prior results).

  Prerequisites: Search space、Sweep job (hyperparameter tuning)

• Machine learning (ML)

  A technique that learns patterns from data to predict on new data; the foundation of AI.

• One-hot encoding

  Converting a categorical variable into 0/1 flag columns, one per category—used for unordered categories (color, region) instead of label encoding (integers, which impose a false order). It increases dimensionality, so for high cardinality consider alternatives (embeddings).

  Prerequisites: Embeddings、ML data preparation (encoding/scaling/imbalance)

• MLOps (SageMaker Pipelines / Model Registry)

  Run the ML lifecycle reproducibly, automatically, and under governance. Pipelines automate process→train→evaluate→register/deploy as a DAG, with an evaluation step + condition as a quality gate. The Model Registry versions models and governs production deployment via approval status. Automate retraining with CI/CD (CodePipeline) or EventBridge.

• AI Gateway (API Management)

  Placed in Azure API Management to centrally authenticate, rate-limit, monitor, and manage token consumption for Microsoft Foundry model access, hiding backend keys. Distinct from Foundry guardrails (model/agent-side safety filters).

  Prerequisites: Azure API Management (APIM)、Token

• Azure OpenAI model deployment and PTU

  In Azure OpenAI you select and deploy a model exposed as an endpoint. Throughput is pay-as-you-go or stabilized via reserved Provisioned Throughput Units (PTU).

  Prerequisites: Azure OpenAI Service

• Generative AI monitoring and tracing

  Monitors model performance and resource consumption via diagnostic settings, records each run via tracing, and collects feedback for continuous improvement.

  Prerequisites: Generative AI

• Conversational Language Understanding (CLU)

  A custom language model extracting intents and entities from user utterances. Add utterances to train, evaluate, deploy, and consume from clients (successor to LUIS).

• Document translation and custom translation

  Translator’s document translation translates files while preserving formatting, and Custom Translator trains/publishes a model adapted to domain-specific terms using parallel data.

  Prerequisites: Azure AI Translator

• Custom Vision training (classification/detection, mAP)

  Choose image classification or object detection, label images, train via transfer learning, evaluate with precision/recall/mAP, then publish and consume.

  Prerequisites: Custom Vision

• DALL-E (image generation)

  An Azure OpenAI model that generates images from text, creating and editing images based on prompts.

  Prerequisites: Azure OpenAI Service

• Document Intelligence prebuilt/layout models

  Prebuilt models that extract common documents (invoices, receipts, IDs) with no training, and a layout model that extracts tables, selection marks, and structure.

  Prerequisites: Azure AI Document Intelligence

• Microsoft Foundry hub and project

  The foundation for generative-AI development. In the hub-based setup, a hub bundles shared resources, connections, security, and cost management, and projects beneath it manage individual workspaces, models, and flows. Since 2025 a hub-less Foundry project (lightweight, standalone) setup is also available.

  Prerequisites: Generative AI

• Image analysis (tags, captions, object detection)

  Image analysis in Azure AI Vision: select visual features to tag images, generate captions, detect objects, and use smart crop to extract key regions.

  Prerequisites: Azure AI Vision

• Key phrase extraction and language detection

  Extracts the main key phrases from text and detects which language it is with a confidence score.

• Model and flow evaluation

  Measures generative-AI output quality via metrics like groundedness, relevance, fluency, and similarity, plus manual evaluation, to improve flows and model choice.

  Prerequisites: Generative AI

• Entity recognition (NER) and PII detection

  Named entity recognition (NER) extracts people, places, organizations, and PII detection finds and masks personally identifiable information.

• OCR (Read)

  Optical character recognition via Azure AI Vision’s Read, extracting printed and handwritten text from images and documents.

  Related: Azure AI Vision

• Prompt shields and harm detection

  Prompt shields detect jailbreak and indirect prompt-injection attacks, and protected-material detection guards against misuse and leakage in generative AI.

  Prerequisites: Generative AI

• Sentiment analysis and opinion mining

  Sentiment analysis scoring text as positive/negative/neutral, and opinion mining extracting opinions per target (aspect).

• Spatial Analysis

  An Azure AI Vision feature detecting people’s presence and movement in video to count people and capture spatial events like distance and entry/exit.

  Prerequisites: Azure AI Vision

• Speech translation and speaker recognition

  Speech translation converting audio to another language’s audio/text in real time, and speaker recognition identifying/verifying a person by voice.

• Azure AI Video Indexer

  Extracts insights—speakers, transcripts, faces, labels, sentiment—from videos or live streams.

  Prerequisites: Index and indexer

• AI observability (tracing/token monitoring)

  Uses Application Insights and Foundry tracing to record per-request latency, token consumption, cost, and errors for continuous quality/performance monitoring.

  Prerequisites: Token

• Fine-tuning and synthetic data operations

  Optimize in order prompting → RAG → fine-tuning (cost/benefit). Fine-tuning hinges on high-quality training data; if insufficient, create synthetic data (managing quality/diversity/bias). Monitor fine-tuned models for overfitting/degradation and manage them dev-to-production via the MLOps lifecycle (evaluate → register → deploy → monitor).

  Prerequisites: MLOps (SageMaker Pipelines / Model Registry)

• Generative AI production monitoring

  Continuously monitors production model outputs to detect quality regressions, groundedness drift, harmful outputs, and data/prompt shifts, triggering alerts and re-evaluation.

  Prerequisites: Generative AI

• Foundry Agent Service (basics)

  A managed service in Microsoft Foundry to build and run agents, combining a model, instructions (system prompt), tools (knowledge, functions), and threads to create goal-oriented AI.

• Foundry evaluation and safety

  Measures generative-AI quality via metrics like groundedness, relevance, and fluency plus manual evaluation, ensuring safety with content filters and risk assessment before deployment.

  Prerequisites: Generative AI

• Model deployment options (Foundry)

  Ways to consume Foundry models: serverless API (pay-as-you-go, ready to use) vs managed compute (deploy to a dedicated endpoint), chosen by requirements from the model catalog.

  Prerequisites: Model catalog

• Custom Vision

  Train on your own images to build custom image classifiers/detectors.

• Face

  Specialized in detecting/analyzing faces in images (some features restricted for Responsible AI).

  Prerequisites: Responsible AI

• Azure AI Translator

  Translates text between languages.

• Backpropagation

  A technique that propagates a neural network's output error backward from the output layer to the input layer, computing each weight's gradient along the way. Paired with gradient descent to update the weights, it's the core algorithm underlying deep-learning training.

  Prerequisites: Gradient descent

• Decision tree

  An interpretable model that repeatedly branches on feature thresholds and predicts at the leaf nodes. Used alone, it tends to overfit the training data and needs depth limits or pruning to control—but it works for both regression and classification.

• AutoML tasks (tabular/vision/NLP)

  Automated ML applies not only to tabular data but also to computer vision (image classification, object detection) and NLP, auto-searching algorithms and preprocessing.

  Prerequisites: Machine learning (ML)

• Command job

  The basic job that runs a single script with a specified environment, compute, and inputs. Pass parameters and data inputs, and diagnose failures via logs.

• Compute instance

  A managed VM for individual development and notebook execution. It can be stopped to save cost and includes a terminal and Jupyter.

• Data assets (uri_file/uri_folder/mltable)

  A reusable, versioned reference to data. Types: uri_file (a single file), uri_folder (a folder), and mltable (a schematized tabular form).

• Fine-tuning job

  A job that further trains a foundation model on custom data to adapt it to a task, proceeding through data prep, base-model selection, training, and evaluation.

  Prerequisites: Foundation model

• Tracing to evaluate a flow

  Recording (tracing) each flow step’s inputs/outputs, latency, and token usage to evaluate and improve quality and performance bottlenecks.

  Prerequisites: Token

• Managed online endpoint

  An endpoint serving real-time low-latency inference where Azure manages infrastructure and autoscale. Kubernetes online uses an existing AKS.

• Primary metric

  The metric a sweep or AutoML optimizes against (e.g., accuracy, AUC), with a direction to maximize or minimize.

  Related: Sweep job (hyperparameter tuning)

• Registries (cross-workspace sharing)

  An org-level catalog to share models, components, environments, and data across multiple workspaces, used to promote from dev to production.

  Prerequisites: Component (reusable step)

• Scoring script (init/run)

  The script implementing inference in a deployment: init() loads the model and run() handles each request. It may be unnecessary for MLflow models.

• Synapse Spark / serverless Spark

  Spark compute to interactively wrangle large data from notebooks, using an attached Synapse Spark pool or Azure ML serverless Spark.

  Prerequisites: Azure Machine Learning

• Traffic split and blue/green

  Place multiple deployments under one endpoint and split traffic by percentage. Roll out a new version gradually and roll back instantly on issues.

• Gradient boosting

  An ensemble technique that sequentially adds weak decision trees, each one trained to reduce the errors left by the trees before it (flagship implementations: XGBoost, LightGBM). A strong default for tabular data, and offered as a built-in algorithm by many managed ML services.

  Prerequisites: Decision tree

• Gradient descent

  An optimization algorithm that minimizes a loss function by repeatedly stepping the parameters opposite the loss gradient, with step size set by the learning rate. Variants include full-batch, per-example stochastic gradient descent (SGD), and mini-batch processing.

• Linear regression

  A foundational supervised-learning model that predicts a continuous target as a weighted linear combination of features. It's easy to interpret and is often the first baseline tried for a regression problem.

• Large language model (LLM)

  A model trained on vast text that generates language by predicting the next token (Transformer-based).

  Prerequisites: Token

• Logistic regression

  A model that passes a linear combination of features through a sigmoid function to output a probability between 0 and 1, performing binary classification. Despite the name, it's a classifier—and like linear regression, it's interpretable and commonly used as a classification baseline.

  Prerequisites: Linear regression

• Microsoft Agent Framework

  An open-source orchestration framework for building complex workflows beyond what a single agent's response can handle, positioned as the successor unifying Semantic Kernel and AutoGen. It supports multi-agent collaboration (role division, handoffs), concurrent sessions across multiple users, and flexible workflow control encompassing both human-in-the-loop (human approval) checkpoints and autonomous execution, all assembled in code. Individual agents are built in the Foundry Agent Service, and this framework composes and runs them together.

  Prerequisites: Foundry Agent Service (basics)

• k-means clustering

  A classic unsupervised method that partitions data into k clusters by iterating assign-to-nearest-centroid then update-centroids. Choose k via the elbow method/silhouette; sensitive to scale, so standardize first. Used for customer segmentation, etc.

  Prerequisites: ML data preparation (encoding/scaling/imbalance)

• Principal component analysis (PCA / dimensionality reduction)

  A dimensionality-reduction method that compresses many correlated features into a few synthetic axes (principal components) maximizing variance—used for visualization, noise reduction, and cutting compute. Unsupervised, with reduced interpretability; t-SNE/UMAP are non-linear alternatives for visualization.

  Prerequisites: Component (reusable step)

• Microsoft Copilot

  A ready-to-use generative AI assistant built into Microsoft 365, Windows, GitHub, and more.

  Prerequisites: Generative AI