AI Copilot (Offline)

The AI Copilot is an offline, on-device assistant that helps you label faster. It can see the current image, answer questions about it, and take labeling actions — all running locally with no cloud calls and no data leaving your machine.

Privacy first. Inference runs entirely on your computer through the same local ONNX Runtime stack as automatic detection. Your images are never uploaded.

What it can do

The copilot turns a chat request into a labeling action and lands the result in the existing review panel, where you accept or reject each suggestion:

• Detect objects and draw bounding boxes (e.g. "detect all the cars").
• Segment objects into masks.
• Caption an image or describe what's in it.
• QA your existing annotations and flag likely mistakes.

Human-in-the-loop by default. The copilot proposes; you dispose. AI-generated boxes and masks arrive as predictions you review, and any change to existing data goes through an approve/deny step.

Availability

Capabilities roll out as their models ship. Object detection and segmentation are available today; richer describe/caption/QA capabilities are being added. Each model shows its status on the AI Models page.

Requirements

The copilot reuses the same runtime as automatic detection, so the setup is the same:

1. ONNX Runtime — installed automatically or manually. See the AI & GPU setup guide.
2. An NVIDIA GPU is optional but recommended for responsive results. CPU works, just slower. See System Requirements.
3. Models — download the models you want to use from the AI Models page inside the app. They are cached locally for offline use.

Getting started

1. Open a project and select an image.
2. Make sure AI Models → Inference Runtime shows ONNX Runtime: Loaded.
3. Open the copilot chat and ask it to do something, e.g. "detect all the people in this image".
4. Review the suggested annotations and accept the ones you want.

If nothing happens when you ask the copilot to detect, the runtime probably isn't loaded — head to the AI & GPU setup guide and check the Inference Runtime panel.