ei-build-custom-ai-labeling-blocks

Help the user author an Edge Impulse custom AI labeling block. An AI labeling block is a Docker container that Studio runs from a project's Data acquisition → AI labeling tab to label a selection of samples with a model or an LLM (single labels, bounding boxes, or audio labels with time ranges). Any language works. The contract is the arguments, environment variables, and Studio API calls the container uses. Custom AI labeling blocks are an Enterprise-only feature.

Before writing a block, check whether a built-in block or an official example already does the job (see Reference repositories). Chaining blocks in one AI labeling action is often better than one large block, for example a zero-shot detector followed by a GPT-4o re-labeling step.

How AI labeling blocks work

An AI labeling block is a transformation block in standalone mode with type: "ai-action" in parameters.json. It receives no input file or directory. Instead Studio passes the IDs of the selected samples in a JSON file, and the block reads and writes each sample through the Studio API with the project API key. There are no required outputs. All changes are applied by API calls from inside the block.

Studio runs the block in two ways:

Required files