Skip to content
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Model stages

Integrate model inference as an observable, reviewable workflow responsibility.

A Model stage makes inference part of the same state machine as human work. It must have a versioned contract, mapped outputs, visible failure behavior, and a human-controlled acceptance route.

Before you make the change

  • Define model owner, version, endpoint or runtime, supported modality, input, output, and ontology mapping.

  • Prepare a calibration cohort and expected failure cases.

  • Decide who handles timeouts, retries, empty output, malformed output, and low-confidence output. A Model stage has one output; it has no separate failure, timeout, or confidence route.

Unitlab AI Models catalog

Treat every public or private model as a versioned production dependency whose mapping and failure behavior are part of workflow design.

Understand the product behavior

  • Model → Annotate → Review: a model proposes labels, an annotator corrects them, and a reviewer verifies the result.

  • Annotate with local assistance → Review: Magic Touch, prompt detection, Find Similar, or tracking helps the annotator within the task.

  • Model → Review with rejection to annotation: model output moves directly to review; rejected work returns to a human correction stage.

  • Annotate → Review → specialist Review: ambiguous or high-risk cases move to a second Review stage configured for specialists.

Workflow save and change impact

Workflows are reusable workspace definitions that can be bound to multiple projects. Saving an active graph is therefore not a cosmetic edit. Before applying a change, Unitlab validates reachability, required edges, terminal stages, and the effect on in-flight items.

Important UX rules:

  • exactly one Project and one Complete stage;

  • Project has no incoming edge and must lead to an entry stage;

  • Review has one Approve and one Reject route;

  • terminal stages have no outgoing routes;

  • every stage is reachable from Project;

  • duplicate outgoing action names are rejected;

  • a stage holding active items cannot be silently deleted;

  • sensitive automated-stage configuration cannot change while occupied without resolving the impact.

When an active workflow change would strand work, Unitlab returns an apply-impact conflict instead of silently moving or losing items.

Model-stage user experience

Entering a Model stage automatically dispatches the configured model. The item shows Processing while inference runs. Predictions are saved as normal annotation history, then the item follows the Model stage’s default edge to another Model stage, Annotate, Review, or a terminal stage. A failed run does not make the item disappear. Depending on the model and the item’s data type, the item either stays in the Model stage, as it does when a run times out or is canceled, or continues to the next stage without the model’s labels. An item that stays shows the status Error in the Task Queue, where a manager can Retry it. An item that continues is marked Model failed in the Task Queue. For your own model, managers also see the reason the run failed.

Geospatial settings

When Data type is Geospatial and the stage uses Unitlab's built-in Prompt Labelling or Find Similar model, the stage shows two more settings:

Setting
What it does
Default

Tile scale

How much ground the model sees at a time. Higher values fit larger objects into view but give small objects less detail: 2× finest detail, 4× cars and small objects, 8× buildings and roofs, 16× very large objects.

Auto, chosen from the image resolution

Auto Refine

Turns building outlines into clean footprints with straight edges and right-angled corners. Outlines that are not box-like, such as roads or rivers, keep their traced shape.

On

A change to either setting applies to model runs that start afterwards; results that were already imported stay as they are. Your own geospatial model does not use these settings: see Bring your own Models. For Auto Refine in the editor, see Geospatial Annotation.

Add a model to production routing

1. Approve the contract

Document inputs, outputs, classes, geometry, confidence, version, and ownership.

2. Map to ontology

Verify every model output has an intentional destination or rejection rule.

3. Place the stage

Connect the stage's input, then connect its single output to the next stage, for example a Review stage where people check the model's labels.

4. Exercise failure modes

Test timeouts, empty output, malformed output, partial success, and unavailable runtime.

5. Review corrected output

Measure human correction by source, class, and model version.

6. Version change safely

Re-run calibration and workflow impact review for material model or mapping updates.

Decisions that affect production

Decision
Production guidance

Pre-label vs autonomous route

Choose according to risk and the required human acceptance gate.

Confidence

Use for prioritization or routing only when calibrated on the relevant domain.

Retry

Inspect remote task state before repeating a mutation.

Ownership

A failed model task must land in a visible queue with a named owner.

Continue the operating flow

  • Monitor Batch Queue and task failures.

  • Inspect systematic correction cohorts.

  • Record model version in release provenance.


Continue with Unitlab: Unitlab’s data annotation platform

Connect model output to QA

Treat successful inference as a proposal that still follows the project's acceptance policy. Route output through human annotation or review where required. Add Consensus when independent human judgments are needed, or Quality Gate when an approved reference is available. A model confidence threshold is not a benchmark-similarity threshold.

Use QA Workflows to connect these checks and their correction paths. Confirm that the model supports the actual source modality and ontology before dispatching work.

Last updated