Azure AI-102 cheat sheet

A single-service resource exposes one endpoint and key for one AI service; a multi-service Azure AI services resource (kind AIServices) exposes one endpoint and key shared by many services.

Each Azure AI services resource issues two keys; regenerating one leaves the other valid, which enables rolling key rotation without downtime.

Keyless authentication uses Microsoft Entra ID with the Cognitive Services User role assigned at resource or resource-group scope.

F0 is a free tier with a capped monthly transaction count and is limited to one F0 resource per service type per subscription; S0 is the standard billed tier.

Setting public network access to Disabled together with a private endpoint forces traffic through the VNet; IP firewall rules are the alternative when public access stays enabled.

Diagnostic settings route request logs and metrics to Log Analytics, Storage, or Event Hubs; platform metrics alone do not capture per-call request content.

Disconnected containers run AI services on-premises but still upload metering data to Azure periodically for billing.

An HTTP 429 response means the RPM/TPM rate limit or quota is exceeded; retry with exponential backoff honoring Retry-After, or raise the tier.

Deployment types: Global Standard routes across regions for highest throughput, Standard pins to the resource region, Provisioned (PTU) reserves throughput for predictable latency, and Batch handles discounted asynchronous jobs.

max_tokens limits only the completion; the context window covers system, prompt, and completion tokens combined.

temperature (0-2) and top_p both control randomness; adjust one, not both, to keep behavior predictable.

A seed value with temperature 0 gives the most reproducible completions, although exact determinism is not guaranteed.

Chat roles are system, user, and assistant; the system message sets persona and rules and precedes user turns.

Content filters score hate, sexual, violence, and self-harm as safe, low, medium, or high and block medium and above by default; custom blocklists match exact terms.

RAG: embed documents into Azure AI Search, retrieve top-k by vector or hybrid similarity, inject the results into the prompt, and return citations.

Embedding models such as text-embedding-3-small and text-embedding-3-large produce vectors compared by cosine similarity; index and query text must use the same model and dimensions.

An agent combines a model deployment, instructions, and tools; a thread stores ordered messages; a run executes the agent against a thread and is polled for status.

Run status requires_action means the model requested one or more tools; the client executes them, submits tool outputs, and the run resumes.

Function calling: declare name, description, and JSON-schema parameters; the model selects the function and returns arguments, and your code performs the call.

File search grounds an agent on uploaded documents via a vector store, chunking and embedding content and returning citations.

Code interpreter runs sandboxed Python for math, data analysis, and chart generation; produced files are downloadable.

The Bing grounding tool returns real-time web answers with URL citations and requires a Bing Search connection on the project.

The OpenAPI tool lets an agent call an external HTTP API described by an OpenAPI 3.0 specification, authenticated through a project connection.

Agent Service runs inside an Azure AI Foundry project; connections store credentials, and managed identity enables keyless authentication to dependencies.

Image Analysis 4.0 visual features include caption, denseCaptions, tags, objects, read, people, and smartCrops; features are requested per call and some paths are asynchronous.

Custom Vision classification assigns a label to the whole image, while object detection returns normalized bounding boxes plus a label per detected object.

Multilabel classification allows several tags on one image; multiclass assigns exactly one tag per image.

Serving predictions requires a separate prediction resource bound to the project; the training resource only trains and evaluates iterations.

Compact domain models export to ONNX, TensorFlow, CoreML, or Docker for offline edge inference, trading some accuracy for small size.

Face verification is a 1:1 check between two faces, identification is a 1:N search within a person group, and both require Limited Access approval.

Read OCR extracts printed and handwritten text as lines and words with bounding boxes; Document Intelligence handles structured forms and fields.

Sentiment analysis returns document-level and sentence-level scores for positive, negative, neutral, and mixed, each with a confidence value.

Opinion mining (aspect-based sentiment) pairs a target with an assessment, for example target food with assessment great.

PII detection returns entities by category and confidence; the redaction policy masks them with a character such as asterisks or with the entity type value.

Custom NER extracts domain-specific entities from labeled utterances; custom text classification assigns one label (single-label) or several labels (multilabel) per document.

A CLU model is built from intents and entities; entity components may be learned, list, regex, or prebuilt, and training modes are standard and advanced.

Custom question answering stores question and answer pairs with alternate questions and supports multi-turn follow-ups and chit-chat.

Summarization offers extractive mode (selects existing source sentences) and abstractive mode (generates new sentences).

Speech capabilities cover speech-to-text, text-to-speech, speech translation, and intent recognition; custom neural voice requires Limited Access approval.

Index field attributes: key identifies the document, searchable enables full-text queries, filterable enables OData filters, sortable enables ordering, facetable enables counts, retrievable returns values.

The enrichment pipeline is data source plus indexer plus optional skillset plus index; the indexer builds the enrichment tree and maps nodes to index fields.

A knowledge store projects enriched content into Azure Storage tables, blobs, or files for downstream analytics.

Built-in skills include OCR, language detection, key phrase extraction, entity recognition, sentiment, image analysis, and text split; custom skills invoke an Azure Function.

Vector search stores embeddings in a collection field and uses an HNSW or exhaustive KNN algorithm configuration for similarity queries.

Hybrid search runs keyword and vector queries together and fuses ranked lists with reciprocal rank fusion; the semantic ranker then re-ranks up to 50 results.

Prebuilt Document Intelligence models include read, layout, invoice, receipt, ID document, business card, and W-2.

Custom Document Intelligence models are template (fixed layout), neural (varying layouts), and composed (multiple models selected by a classifier).