Skip to content

User guide

For the people who annotate, review and run projects. Administrators installing the platform should read INSTALL.md first.

Role Can
Annotator Take tasks from the queue, annotate, save drafts, submit, comment
Reviewer Everything an annotator can, plus approve or reject submitted work, correct it, run bulk actions
Owner Everything, plus project settings, label schema, members, imports, exports, snapshots, pre-labelling and webhooks
Viewer Read only

An owner can limit an annotator, reviewer or viewer to some folders of the project (Settings → Members → Folders). Items outside those folders do not exist for that person.

Open a project and press Start annotating. The queue hands you the most urgent open task and locks it to you; Release gives it back. Opening an item from the grid is only browsing and takes no task.

  • Save draft keeps your work without sending it anywhere.
  • Submit (Ctrl/⌘ + Enter) checks the result against the label schema and sends it to review. A missing required field or a wrong value is listed before anything is sent.
  • N / P move to the next / previous item; Shift + C copies the shapes of the previous saved item.
  • Every save is a new version. Nothing you saved is ever overwritten.
Item How you annotate
Image Pick a class, then a tool: Select (V), Hand (H), Box (B), Rotated box (R), Polygon (G), Polyline (L), Point (P), Smart polygon (S, a model draws the outline from a click or a box; see below), Keypoints (J), Brush (K) and Eraser (E) for masks, Superpixels (X), Measure (M). Zoom with the wheel or + / −, fit with 0. Move the view by dragging empty image with Select, anywhere with Hand, with the middle mouse button, or by holding Space (you return to your tool on release). Delete or Backspace removes the selected shape. Very large images load tile by tile.
Video Boxes as keyframes on a track; frames in between are interpolated. Step frame by frame and mark where an object leaves the view.
Text Select text to mark a span of the active class; link two spans with Link (R) to make a relation.
PDF Boxes and text spans per page; Read text (OCR) for scanned pages.
Audio Drag across the waveform to mark a segment; type the speaker and the transcript in the list below. K plays and pauses; Delete removes the selected segment.
Time series Each channel has its own lane. Drag across the plot to mark an interval, and choose whether it covers every channel or only the ones shown.
LLM evaluation Read the conversation and the candidate responses. Rank them (move up / down), or pick the better of two. Rate each response, each turn or the whole conversation on the scale.

Smart polygon. Pick the model under Smart polygon in the sidebar (your administrator registers it, for example the bundled SAM2), press S, then click an object or drag a box around it; the outline arrives as an ordinary polygon you can edit. The first click on an image takes about a second while the model reads it, later clicks are near-instant. A class only takes the shapes its schema allows: if the active class is box-only, the canvas says so and nothing is drawn, so pick a class that allows polygons first. The same holds for every drawing tool, and a submit that still breaks a schema rule lists what is wrong.

Context. When the item has companion views (a caption, a second image, the source text), they appear under Context in the right panel. Answer questions about them, such as whether the caption matches, in the Classification fields.

Attributes. Select a shape to edit its attributes in the right panel.

Start reviewing takes submitted work from the review queue.

  • Approve accepts the version.
  • Reject sends it back to the annotator with your comment.
  • You can correct the result yourself before approving; the correction is a new version in your name.

Bulk actions on the item grid (approve, return to queue, assign, tag) apply to many items at once. Items they do not fit are listed with a reason.

Comment on an item, or on one shape of it. Mention someone with @their@email. Mentions, replies to your comments and review verdicts appear under the bell. If your administrator has set up e-mail, they also come by e-mail; turn that off on your Security page.

  1. Storage. An administrator registers the connector (Connectors page). Check confirms the platform can reach it and that the browser will be allowed to load from it.
  2. Project. Create it, pick the source (and, optionally, a different result) connector, and set the folder (prefix) and file pattern (glob). In Settings you can also make same-name files such as photo.txt beside photo.jpg a context view instead of an item.
  3. Label schema. Classes with their tools, colours, hotkeys and attributes; rating classes need a scale. Each save is a new version, and old annotations keep pointing at the version they were made with.
  4. Workflow. Review on or off, consensus (several annotators per item), gold items for accuracy checks, priorities and deadlines.
  5. Items. Scan the source, upload files from the browser, or import existing annotations (COCO, YOLO, VOC, CVAT, Label Studio). Coming from Azure Machine Learning data labeling: MIGRATE_AZURE_ML.md.
  6. Pre-labelling. Run a registered model over new items. Its suggestions arrive as drafts that people correct, and they never replace human work.
  7. Members. Add people with a role, optionally limited to folders.
  8. Results. Every saved version is written to the result storage as it happens. Snapshots freeze a dataset (optionally split into train / val / test). Exports produce COCO, YOLO, YOLO-pose, spaCy, CoNLL, LLM preference pairs, segments (with RTTM) or the native format.
  9. Quality. The dashboard shows progress and throughput; the Quality panel shows agreement between annotators and accuracy on gold items.
  10. Integrations. Webhooks for your own systems (signed JSON), Slack or Microsoft Teams messages, and retraining events for ML pipelines (MLflow, Databricks, Azure ML).
  • API keys (Settings → API keys) for scripts and CI; a service account for anything that is not a person.
  • Python SDK and CLI: pip install annotide, then for example annotide export <project> --format coco -o dataset.zip (sdk/README.md).
  • AI agents: annotide mcp lets an MCP client such as Claude pre-label a project with a service account’s key (sdk/README.md).