Glossary
This page defines the words that the console and these docs use for workflows, quality and review. Each entry is one or two sentences. The guides explain each part in full.
Workflows
Section titled “Workflows”Workflow
Section titled “Workflow”A durable process installed on a graph. It has a name, a trigger and its settings. See Workflows.
Trigger
Section titled “Trigger”What starts a workflow: on change, on upload, on schedule, on stream or on review answered. Today a workflow has one trigger: a watch (on change) or a content trigger (on upload). On schedule, on stream and on review answered are planned.
Template
Section titled “Template”A ready-made, versioned starting point for a workflow. The API calls it a starter workflow. Use template creates a workflow from it. See Starter workflows.
One item of work. A run groups its continuations and its retries.
Attempt
Section titled “Attempt”One try of a run. The console shows attempts only inside the details of a run.
A checkpointed unit inside a run. When an attempt runs again, a finished step returns its saved result.
AI steps
Section titled “AI steps”The steps that use a model or a person: extract, classify, decide, embed and review. A workflow’s code makes these calls through the activities.
Agent step
Section titled “Agent step”A model that uses tools in a bounded loop. Agent steps are planned.
Quality
Section titled “Quality”Dataset
Section titled “Dataset”Example inputs with their expected outputs. Datasets are coming to Evals.
A run on chosen inputs before a workflow goes live. A test writes nothing. Tests are coming.
Experiment
Section titled “Experiment”Two versions compared on one dataset. The console calls model trials experiments.
A judge model grades a sample of the model calls of a use. People can agree with a check or correct it. See Check model quality.
The confidence below which a workflow asks a person instead of writing. A result at or above the gate is written.
Calibration
Section titled “Calibration”People label a sample of results. LBB uses the labels to measure a gate and set it.
Review
Section titled “Review”Review item
Section titled “Review item”A question for a person: approve, correct, choose, grade or answer. Its result is an answer.
Review
Section titled “Review”The way a workflow hands a decision to a person, and the inbox of review items. Today the inbox is the Suggestions tab of Knowledge graph.
What a model call is for: ask, route, extract, decide, embed and more. Models picks the model for each use.
Engine words
Section titled “Engine words”The console shows these words only under a Details disclosure. You meet them in the workflow engine’s routes and in its logs.
| Word | Meaning |
|---|---|
| instance | An engine identity that receives messages. |
| turn | One attempt of handling a message. |
| shard | One of up to 16 identities that a workflow’s work is routed to. |
| lease | A worker’s claim on a turn. |
| inbox | The queued messages of an instance. |
| origin | What caused a message or a review item. |