AI-300 key terms

A central Azure resource that groups compute, data connections, environments, models, and jobs so a team can build, train, and track machine learning work in one governed place.

A managed pool of virtual machines in Azure Machine Learning that runs training jobs or batch scoring, automatically scaling nodes up for demand and down to zero when idle.

A saved, credential-secured connection from a workspace to storage such as Azure Blob Storage or Azure Data Lake, letting jobs read and write files without hard-coded secrets or paths.

A versioned recipe of a base container image plus packages and environment variables that pins the software runtime for training or scoring, so runs behave the same everywhere.

A named, versioned entry in the workspace model registry holding model files plus metadata such as metrics and tags, so deployments always reference a fixed, auditable artifact.

A reusable, self-contained pipeline step with its own code, environment, inputs, and outputs that can be shared across pipelines and rerun in isolation during debugging.

An open standard logging interface that records parameters, metrics, and artifacts for each run; Azure Machine Learning can store those runs and show them beside native jobs.

A network interface that gives a workspace or its dependencies an address inside a virtual network, so traffic stays private and public internet access can be disabled.

A hosted HTTPS service that answers individual requests in real time by routing them to one or more model deployments, with authentication, scaling, and logs handled by the platform.

A deployment target that runs a scoring job over a large file or folder on demand or on a schedule and writes predictions to storage, instead of replying to single requests.

The percentage of requests an online endpoint sends to each of its deployments; shifting it gradually from old to new lets a team canary-test a model before full rollout.

A measurable change in the distribution of live input features compared with the training data, which can degrade predictions even when the model itself has not changed.

A rule such as a schedule, a new data landing event, or a drift alert that starts a training pipeline automatically so a deployed model is refreshed.

Automated build and release stages that test code, rebuild environments, train or evaluate a model, and promote artifacts between dev, test, and production environments.

A workspace view that combines error analysis, fairness assessment, interpretability, and causal analysis so reviewers can judge a model's behavior before and after deployment.

A shared Azure resource that supplies networking, identity, and connections for several AI projects, giving administrators one place to govern generative AI resources and cost.

A workspace inside a hub that holds a team's model deployments, prompt flows, datasets, indexes, and evaluation results for one application or use case.

A reserved block of model capacity billed hourly that guarantees a fixed level of processing for a deployment, giving steadier latency than pay-per-token billing at high volume.

A stored, credential-holding reference from a project to an external service such as Azure OpenAI, Azure AI Search, or storage, so flows use it without embedding keys.

A development tool that describes a generative application as a graph of nodes for prompts, Python code, and tools, with defined inputs, outputs, and inspectable runs.

A prompt flow published as a hosted online endpoint that accepts requests, runs the flow, and returns responses while the platform handles authentication, scaling, and monitoring.

A searchable collection of chunked documents and their vector embeddings that an application queries at request time so generated answers are grounded in your own content.

A policy layer that classifies prompts and completions into harm categories such as violence or hate and blocks or annotates content above configured severity thresholds.

A curated set of test inputs paired with expected answers or ground-truth context, used to score an application the same way each time versions are compared.

An evaluator score estimating whether the claims in a generated answer are supported by the retrieved context, flagging responses that add unsupported or invented information.

An evaluator score estimating how directly a generated answer addresses the user's question, regardless of whether its wording matches a reference answer.

An evaluation method in which a capable model scores another model's output against a written rubric, making it practical to judge qualities like coherence at scale.

A Python function you write that applies your own rules or metrics to flow outputs and runs alongside built-in evaluators in the same evaluation job.

Recording every step of a flow or application run, including inputs, outputs, latency, and token counts, so slow steps and failures can be found and fixed.

Scoring a sample of live production requests with evaluators and sending the scores to monitoring, so quality drops are caught before users report them.

Structured adversarial testing that probes a generative application with attack prompts to expose harmful, unsafe, or policy-violating outputs before release.

The instruction block placed at the start of a conversation that sets the model's role, tone, limits, and output format, shaping every later turn.

Further training of a base model on your own labeled examples so it learns a task, style, or output format more reliably than instructions alone can achieve.

Training a smaller, cheaper model to imitate the outputs of a larger, stronger model, aiming to keep much of the quality at lower cost and latency.

A layer that inspects each request and sends it to the cheapest or fastest model that can handle it, balancing quality against cost across a model portfolio.

Storing earlier responses keyed by the meaning of the request so a new but similar question returns the saved answer instead of calling the model again.

Splitting source documents into smaller, often overlapping passages before embedding, so retrieval returns focused context that fits inside the model's token limit.

The combined limit on prompt and completion tokens for a request, which caps answer length, cost, and how much retrieved context can be supplied.