Glossary

The core vocabulary of Pradra Studio, in plain terms.

The building blocks

AI Solution — A business problem represented in the platform; a container that groups related capabilities. See Solutions.

AI Capability — One reusable, contract-bound unit of intelligence inside a solution: a prediction, a generation, or an agent. See Capabilities.

Contract — A capability's declared inputs (field names and types) and outputs. Every prediction request is validated against it.

Acceptance criteria — The quality rules a trained model must satisfy before it can go to production (e.g. "accuracy ≥ 0.85"). Set once per capability, enforced on every candidate, forever. See Acceptance criteria.

Data

Dataset — A validated, versioned collection of training/evaluation data. See Datasets.

Feature set — A versioned, model-ready view of a dataset: the columns, derived values, and target used for training and prediction — identical at train and serve time. See Feature sets.

Connector — A registered link to an external data source (Postgres, MySQL, S3) that syncs data in as new dataset versions. See Connectors.

Training-serving skew — The bug where a model sees differently-prepared data in production than in training. Prevented structurally by feature sets.

Models

Experiment — The immutable record of one training run: parameters, metrics, logs, artifacts, lineage. Produces one candidate model. See Experiments.

Model — A trained, versioned artifact that backs one capability at a time.

Model registry — The governance layer that moves a model through Registered → Staging → Production → Archived, only ever promoting models that pass acceptance.

Promotion — Making a published model the capability's production version.

Rollback — Pointing production back at a previous model version. A pointer switch — instant, never a retrain.

Serving

Deployment — A running instance of a capability's production model, reachable as a stable endpoint. See Deployments.

Prediction history — The audited record of every prediction: inputs, outputs, latency — and for generative calls, tokens and cost.

Batch prediction — Scoring many records at once as an asynchronous job, with a downloadable result. See Batch predictions.

API token — A long-lived credential (prefix aif_…) your application uses to call the API. See API tokens.

Operations

Drift — A change over time in the data arriving in production (data drift) or in the model's own predictions (model drift). The early warning that the model's world has moved. See Drift & retraining.

PSI — The statistical score used to measure drift, read in bands (low / medium / high).

Retraining policy — The per-capability rules for automatic retraining: triggers (manual, schedule, drift), cooldowns, and approval mode.

Budget — A monthly cap on generative spend (USD and/or tokens) at the organization or capability level. A hard stop, not a suggestion. See Budgets & costs.

Generative AI

Model adapter — A registered connection to an LLM provider or self-hosted model; capabilities pin an exact model identifier through it.

Prompt version — One immutable revision of a generative capability's instructions. Activation is gated by evaluation. See Prompts.

Golden set — Curated input → expected-output examples that anchor generative and agent evaluation. See Generative evaluation.

Knowledge source — A versioned, workspace-scoped store of your documents, chunked and embedded for retrieval-augmented generation. See Knowledge sources.

Fine-tuning — Training the model itself on your examples, self-hosted or via a provider (with explicit consent). See Fine-tuning.

Automation

Workflow — A validated graph of steps — capabilities, conditions, transforms, approvals — executed as durable, traceable runs. See Workflows.

Agent — A capability that pursues a goal by planning and calling tools in a loop, inside hard guardrails, with every step traced. See Agents.

Agent tool — A registered, contract-bound action an agent may be granted; classified read_only, reversible_write, or consequential (which always requires human approval). See Agent tools & approvals.

Platform

Organization — The top-level tenant. Owns workspaces, users, budgets, and its subscription. See Organizations & workspaces.

Workspace — The isolation and collaboration boundary inside an organization. Everything you build lives in one.

Secret — A centrally stored, write-only, typed credential the platform uses to reach your systems. See Secrets.

Plan / entitlement — The subscription plan's caps on how many resources an organization may create. See Subscription & plans.

Audit trail — The record of who did what, and when — with credentials never written into it. See Governance & audit.