# Glossary

Source: [https://docs.aventail.co.uk/docs/introduction/glossary](https://docs.aventail.co.uk/docs/introduction/glossary)

| Term | Meaning |
| --- | --- |
| **Agent** | Link's background runtime on a device. |
| **API base URL** | The address a client uses for inference, usually ending in `/v1`. |
| **Canary** | A small initial group used to evaluate a rollout before expanding it. |
| **Control** | Aventail's hosted interface for managing devices and models. |
| **Deployment** | Downloading and preparing a model on a device; distinct from serving it. |
| **Device** | A registered machine running Link. |
| **Endpoint** | An address and path used for an API request. |
| **Fleet** | An organisation's managed group of devices. |
| **GGUF** | A model file format used by the language-model runtime. |
| **Inference** | Running a model on an input to obtain a result. |
| **Aventail Link (Link)** | Software installed on the machine that runs your models. Version 1.x is documented as Locai Link (Legacy Aventail). |
| **llama.cpp** | An inference engine used for compatible language models. |
| **Model ID** | The exact identifier returned by the serving endpoint's `/v1/models` response. |
| **Model library** | Models available in Control, including owned and globally shared models. |
| **Overlay / revision** | A saved set of fleet configuration values. |
| **Registration key** | A credential used to enrol a headless device. Its expiry is set in Control. |
| **Ring** | A group of devices in a staged rollout. |
| **Serving** | Making a deployed model available to accept requests. |
| **Telemetry** | Measurements and operational state reported by a device. |
| **TensorFlow Lite / TFLite** | A format and runtime used by supported classification pipelines. |
| **WebGPU** | Browser graphics/compute technology used by Workspace's **This device** models. |
| **Workspace** | Aventail's chat and document interface. An organisation workspace in Control is a separate account context. |