AI-200 flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Azure AI Foundry project

A container inside an Azure AI Foundry resource that groups connections, model deployments, agents, and assets for a team, inheriting the parent resource's security, quota, and networking settings.

Global Standard deployment

A deployment type that routes requests to whichever region has available capacity; it offers the lowest cost per token, but prompts and completions may be processed outside the deployment region.

Provisioned Throughput Unit

Pre-purchased model capacity measured in PTUs that guarantees a fixed throughput and latency profile, billed hourly, and suited to steady high-volume production workloads.

Cognitive Services OpenAI User role

An Entra ID role granting permission to call model inference endpoints on an Azure OpenAI or AI Foundry resource; it does not permit creating deployments or managing keys.

Content filter severity levels

Azure AI Foundry grades harm categories hate, violence, sexual, and self-harm as safe, low, medium, or high; default filters block medium and high severity on both prompts and completions.

Managed identity for AI resource

An Entra ID identity assigned to the AI resource so it can reach Storage, Search, or Key Vault without stored secrets; the target resources must grant that identity the appropriate RBAC roles.

Temperature parameter

Controls randomness of generated text: values near 0 make output deterministic and factual, higher values increase diversity; adjust temperature or top_p, not both at once.

Retrieval-augmented generation

A pattern where relevant documents are retrieved from an index and injected into the prompt as grounding context, reducing hallucination and letting the model answer over private data.

Groundedness evaluation metric

Measures whether the claims in a generated answer are supported by the supplied context; a low score signals hallucination even when relevance and fluency scores are high.

Structured outputs json_schema

A response_format option that forces the completion to conform to a supplied JSON schema, returning valid parseable JSON and removing the need for retry-and-parse logic in application code.

Function calling flow

The model returns a tool call with JSON arguments instead of text; the application executes the function, appends the result as a tool message, then calls the model again for the final answer.

Embedding vector dimensions

The length of an embedding vector; newer text-embedding-3 models accept a dimensions parameter to shorten output, and stored index dimensions must match query dimensions exactly.

Foundry Agent Service thread

Server-side conversation state holding an agent's message history; you append user messages and create runs against the thread instead of resending the full chat history each time.

File search tool

An agent tool that ingests uploaded documents into a vector store, chunks and embeds them automatically, and performs retrieval so the agent can answer questions about those files.

Code interpreter tool

A sandboxed Python execution tool that lets an agent perform calculations, generate charts, and process uploaded files, returning textual output plus generated images or data files.

Connected agents

A capability where a primary agent invokes other agents as tools, enabling multi-agent orchestration in which each sub-agent has its own instructions, model, and toolset.

Agent run status

A run progresses through queued and in_progress to terminal states completed, failed, cancelled, or requires_action; requires_action means you must submit tool outputs before the run continues.

Custom Vision classification vs detection

Classification assigns labels to a whole image; object detection also returns bounding boxes per object and is required whenever object location or count matters.

Image analysis captions

A feature returning a one-sentence natural-language description of an image; dense captions return separate descriptions for multiple regions of the image, each with a bounding box.

OCR read model

Extracts printed and handwritten text as lines and words with bounding polygons; use it for plain text, and Document Intelligence layout when table or paragraph structure is needed.

Custom Vision iteration

A trained version of a Custom Vision project; you review its precision, recall, and mAP scores, compare iterations, then publish the chosen one to a prediction resource endpoint.

Spatial analysis

A video capability that analyzes people in camera streams to produce counts, zone entries, and line crossings, emitting JSON events rather than returning processed video.

CLU vs custom question answering

CLU extracts intents and entities from utterances for command-style bots; custom question answering instead matches user questions to a curated FAQ knowledge base and returns the best answer.

PII detection

An Azure AI Language feature that identifies personal data such as names, phone numbers, and identifiers in text and optionally redacts it using a configurable masking policy.

Sentiment analysis confidence scores

Returns positive, neutral, and negative scores per document and per sentence; the highest score determines the label, and opinion mining adds aspect-level sentiment targets.

Speech intent recognition

Combines speech-to-text with a CLU or LUIS model so a spoken utterance returns transcribed text plus the top intent and extracted entities in a single service call.

Precision and recall tradeoff

Precision is true positives divided by predicted positives; recall is true positives divided by actual positives; F1 is their harmonic mean, used to compare threshold settings.

Search index field attributes

searchable enables full-text matching, filterable enables OData filters, sortable enables ordering, facetable enables counts, retrievable returns the field, and key uniquely identifies documents.

Semantic ranker

Reranks the top keyword or fused results using a language model to improve relevance for natural-language queries; it supplements the base ranking rather than replacing it.

Hybrid search with RRF

Runs keyword and vector queries in parallel and merges the ranked lists using reciprocal rank fusion, which combines by rank position and requires no score normalization between methods.

Integrated vectorization

An indexer-driven pattern where a skillset chunks documents and calls an embedding deployment, populating the vector field during indexing instead of embedding on the client.

Knowledge store

An optional projection of enriched documents from a skillset into Azure Storage as tables, JSON objects, or files, enabling downstream analytics, dashboards, and visualization.