Minifield builds small models for specific software features. Give a model your product’s vocabulary, data, and supported actions, then run inference locally in your application.
These docs cover the interfaces you embed, the runtime that executes your model, and the training worker that prepares, trains, and evaluates it.
Find your starting point
Extract typed fields from free text in a React textbox.
RuntimeRun LFM2 inference on CPU, WebGPU, and native Metal.
CommandTurn requests into proposals users can review and approve.
ObserverTrack usage and inspect findings in local AI tool logs.
TrainingTrain and evaluate a model around your product’s supported scope.
How the pieces fit
Your application supplies the product context and enforces permissions. MagicBox presents extracted fields. Command presents action proposals and asks the host to execute the selected action. Runtime handles local model inference.
Training turns authored cases into reproducible inputs, runs optimization and evaluation, and exports model assets. Runtime loads those assets on the user’s device.
Read the integration guide for the responsibilities at each interface.
Begin with one feature
Choose a task you can describe and test: extracting an email address, selecting a supported filter, or changing a document property. Define its inputs, outputs, and permitted actions before selecting a model.
Start with MagicBox’s React integration to connect an extraction callback. For local inference, follow the runtime guide.