AI-200 cheat sheet

Deployment types: Global Standard (lowest cost, data may leave the region), Data Zone Standard (processed within a geography such as EU or US), Standard (single region), Provisioned (reserved PTUs billed hourly), Batch (async, roughly 50% cheaper, 24-hour target turnaround).

Quota is allocated per subscription, per region, per model in tokens-per-minute and requests-per-minute; calls beyond quota return HTTP 429 with a Retry-After header.

Role choices: Cognitive Services OpenAI User = inference calls; Cognitive Services Contributor = manage resource and deployments; Azure AI Developer = build and deploy in Foundry projects; Owner/Contributor = RBAC and key management.

Prefer Microsoft Entra ID authentication with a managed identity, and disable local (key) authentication on the resource so callers must present Entra tokens.

Private networking requires a private endpoint for the AI resource plus public network access disabled; indexers additionally need a shared private link from Azure AI Search to the target service.

Content filters are configured per deployment and grade hate, violence, sexual, and self-harm on safe/low/medium/high; prompt shields block jailbreak (user prompt) and indirect (document-embedded) attacks.

Diagnostic settings categories are Audit, RequestResponse, and Trace sent to Log Analytics or Storage; platform metrics cover token counts, latency, and throttling.

Cost drivers: input and output tokens, hourly PTU reservation, vision transactions per 1,000, skillset enrichment and indexer runtime, plus storage for indexes and knowledge stores.

Chat request body: messages with system, user, assistant, or tool roles, plus temperature, top_p, max_tokens, stop, seed, frequency_penalty, presence_penalty, response_format, and stream.

prompt_tokens plus completion_tokens must fit the model context window; max_tokens caps only the generated completion, never the prompt.

Adjust temperature or top_p, not both: 0-0.3 for extraction/classification, 0.7 or higher for creative variation; fixed seed with temperature 0 gives near-deterministic output.

RAG pipeline: chunk (512-1,024 tokens with overlap), embed, index in Azure AI Search, retrieve top-k with hybrid plus semantic ranker, inject with citations, and instruct the model to answer only from the supplied context.

finish_reason values: stop (natural end), length (max_tokens reached), content_filter (blocked), tool_calls (function calling); inspect prompt_filter_results when a prompt is blocked.

response_format with json_schema and strict:true forces schema-conformant JSON; function calling instead returns JSON arguments that your code executes before a follow-up completion.

Evaluators: groundedness (claims supported by context), relevance (addresses the question), coherence, fluency, and similarity against a ground-truth answer; compare the same dataset to a baseline.

Fine-tuning shapes style, format, and tone but does not reliably add factual knowledge; use RAG or grounding for current or private facts.

Agent anatomy: a model deployment plus instructions, tools, and optional knowledge (vector store); the definition is stored server-side in a Foundry project and referenced by an agent ID.

Invocation flow: create thread, add a user message, create a run, then poll or stream until the run reaches a terminal status (completed, failed, cancelled, requires_action).

requires_action means the model requested tool output: submit the tool results to the run and the run resumes toward completion.

Built-in tools: file search (vector store retrieval), code interpreter (sandboxed Python, charts, file output), Bing grounding, Azure Functions, OpenAPI-specified APIs, MCP servers, and connected agents.

File search manages chunking and embedding automatically; quotas apply to individual file size, number of files, and total vector store capacity.

Connected agents let a primary agent invoke other agents as tools, which implements sequential, concurrent, and handoff multi-agent patterns.

Use explicit code-defined orchestration when the workflow steps are fixed; allow autonomous tool selection when the path depends on user input.

Emit agent traces and tool-call spans to Application Insights and log token usage per run for cost and quality monitoring.

Image Analysis 4.0 returns only the features explicitly requested: caption, denseCaptions, tags, objects, people, read, smartCrops, detect.

Use Read OCR for plain text extraction (synchronous for images, asynchronous for documents); use Document Intelligence prebuilt-layout when tables, selection marks, or structure matter.

Custom Vision project types: classification (multiclass = exactly one label, multilabel = several) and object detection (bounding boxes plus labels).

Object detection needs far more labeled images than classification; keep classes balanced and start from roughly 15-30 images per class.

Custom Vision metrics are precision, recall, and mAP; tune the probability threshold at prediction time and the overlap threshold when matching predicted boxes.

Publish the chosen iteration to a prediction resource; compact domains allow export to ONNX, TensorFlow, or CoreML for edge inference.

Face detection returns a faceId plus attributes; identification and verification need a person group or large person group; liveness detection blocks photo and mask spoofing.

Capability selection rule: classification for a label per image, detection for location or count, OCR for text, spatial analysis for people counts and zone or line-crossing events in video.

Azure AI Language capabilities: sentiment with opinion mining, key phrase extraction, named entity recognition, PII detection and redaction, entity linking, language detection, and extractive or abstractive summarization.

CLU extracts intents and entities from command-style utterances; custom question answering matches questions to a curated knowledge base and returns an answer with confidence and source.

CLU training requires tagged utterances per intent; evaluation reports precision, recall, and F1 per intent and entity, and a trained model must be deployed to an endpoint before it can be called.

Custom question answering sources are imported FAQ pages or URLs, documents, and manual pairs; alternate questions improve matching and multi-turn follow-up prompts handle clarifications.

PII redaction policies: mask with a character (default asterisk), mask with the entity type, or no mask; a confidence threshold filters low-confidence detections.

Speech-to-text with intent recognition combines transcription and CLU in one call, returning text, top intent, and extracted entities.

Translator provides text translation, transliteration, and language detection; Custom Translator adds domain terminology using parallel documents.

Tune classifiers with the confidence threshold: raising it increases precision and lowers recall; F1 is the harmonic mean used to compare configurations.

Index field attributes: key (Edm.String, unique, required), searchable (full-text), filterable, sortable, facetable, retrievable; only retrievable fields appear in results.

Only simple string fields can be searchable; complex types must be nested as Edm.ComplexType, and collections use Collection(Edm.String) or Collection(Edm.ComplexType).

Ranking: BM25 by default; scoring profiles add field weights and functions (magnitude, freshness, distance, tag); the semantic ranker reranks the top ~50 results for natural-language queries.

Vector search options: HNSW (approximate, fast, tuned via m, efConstruction, efSearch) versus exhaustive KNN (exact, slower); vector dimensions must match the embedding model output.

Hybrid search runs keyword and vector queries together and merges them with reciprocal rank fusion, so scores from the two methods need no normalization.

Indexer pipeline: data source, indexer with schedule and change detection, skillset enrichment, target index, plus optional knowledge store projections.

Common skills: OCR, split (chunking with overlap), merge, language detection, entity and key phrase extraction, embedding, and custom Web API skills.

Document Intelligence models: prebuilt-read (text), prebuilt-layout (structure and tables), prebuilt-invoice/receipt/idDocument (fields), custom template (fixed layout, 5+ labeled samples), custom neural (varying layouts).

A knowledge store projects enriched content into Azure Storage as tables, JSON objects, or file projections for downstream analytics and visualization.