Definitions used across the SoAI documentation. Follow the linked guide when you need to configure or operate a feature.
Product and access terms
| Term | Meaning |
|---|---|
| Access action | An ACL permission that gates a WebUI surface or a management API operation. The policy assigns actions to roles rather than to individual accounts, so changing one entry changes it for every account holding that role, and administrators edit it in Settings > ACL. See Access control for the permission table. |
| Anthropic-compatible API | The Anthropic-shaped Messages surface at /v1/messages and /v1/messages/count_tokens. It shares the /v1 prefix and the API-key policy with the OpenAI-compatible routes while using Anthropic request, response, and streaming shapes. See Anthropic-compatible Messages API. |
| API key | A bearer credential for the compatible API clients on /v1. Complete first-run setup before calling those routes; after setup the surface stays open while no keys exist, and creating the first key turns API-key authentication on for every /v1 route until the last key is removed. Administrators manage keys in Settings > OpenAI. See Authentication and API keys. |
| Backend | The process or service that performs inference behind a plugin, either a managed local engine or a remote provider. A plugin can be installed and enabled while its backend is absent, installing, stopped, starting, or running. See Install and manage plugins. |
| Deployment | One independently managed SoAI OS control plane with one licensing identity, which is how SoAI OS entitlements are counted. The Core license sets no deployment limit for Personal Use. The licensing documentation explains how replicas, standbys, and backups are counted. |
| OpenAI-compatible API | The OpenAI-shaped routes under /v1, covering the model catalog, chat completions, completions, embeddings, images, audio, Responses, and files. The Anthropic-compatible Messages routes share the same prefix and key policy. See HTTP API reference for the current route inventory. |
| SoAI Core | The Core application: the browser-based management UI, compatible APIs, routing, plugins, and associated first-party Core source. See SoAI products and editions for the edition boundary. |
| SoAI OS | A separate private appliance edition built on Core, with its own launcher, host-management features, release payload, and license. See the SoAI OS overview. |
| WebUI | The signed-in browser interface used to manage SoAI. To enter Settings, select the Settings button at the right of the header, then select Settings from its menu. Documentation names the tab that follows, such as Backup. See WebUI architecture for the interface boundary. |
| Workspace | The directory tree assigned to a signed-in account and the only part of the host that Files, its API, and workspace-scoped tools can reach. Paths shown in Files are relative to it. See File explorer. |
Models and plugins
| Term | Meaning |
|---|---|
| Circuit breaker | A resilience guard that limits dispatch to a repeatedly failing plugin so that routing can protect unrelated work. See Resilience and self-healing for recovery behavior. |
| Clone | A second independent instance of a plugin that supports cloning, with its own name, its own configuration file, and optionally its own copy of the model files. See Install and manage plugins. |
| Model ID | The value selected in the WebUI or supplied in an API request's model field. SoAI stores each catalog entry under a universal ID of the form <plugin>-<source model id>-<hash> and resolves a plugin/model name or an assigned alias to that ID before routing. The Models catalog explains the entries that SoAI can route. |
| Plugin | A .soaiplugin package that connects SoAI to an inference backend or provider. SoAI runs a loaded plugin through a dedicated worker process. See Plugin architecture for the runtime boundary and the Plugin SDK reference for the public contract. |
| Plugin Guardian | The service that watches plugin runtime health, pings idle plugins, and performs bounded recovery. It is separate from the dispatch circuit breaker, and either one can make a plugin unavailable for a different reason. See Resilience and self-healing. |
| Plugin state | The single label a Plugins card shows for a plugin, such as Ready, Idle, Backend not installed, Error, or Quarantined. It collapses conditions that change independently, including package compatibility, backend installation, worker lifecycle, and Guardian health, so read the card detail before restarting or reinstalling a backend. Plugin architecture lists those axes, and Install and manage plugins covers state-specific guidance. |
| Quarantined | A plugin state that blocks normal work after recovery escalation or an integrity problem. Inspect the plugin card, then follow Make a quarantined plugin available again. |
| Virtual model | One routable model name backed by at least two constituent models. It can use load balancing to distribute requests or failover to try alternatives after a failure. Manage it through Models > Virtual models; the control requires Model routing (MODEL_ROUTING_ADMIN). See Virtual models, routing, and failover. |
Work, knowledge, and integrations
| Term | Meaning |
|---|---|
| Knowledge or RAG collection | A conversation-specific set of indexed sources used to retrieve relevant passages during a later response. In Chat, select Add attachments > Knowledge to add sources and Chat configuration > Knowledge settings to configure retrieval. See Knowledge and document collections (RAG) and Attachments and file context. |
| MCP | Model Context Protocol. In host mode, SoAI connects to an external MCP server. In server mode, SoAI exposes selected capabilities to authenticated MCP clients at /mcp over Streamable HTTP. Administrators manage connections in Settings > MCP. See the MCP overview. |
| Skill | A reusable SKILL.md instruction package, optionally with supporting files. The account owns its library, and each ordinary chat selects the skills it can use. See Skills for automatic loading and explicit name requests. |
| SoAI Bench | The GPU benchmark surfaced on the Hardware page. Its OpenCL preflight is separate from normal GPU monitoring and can report a device as unsupported for benchmarking. See the SoAI Bench overview. |
| Task | A trackable operation such as a model download, plugin installation, automation run, or maintenance action. Open the header bell and select Tasks to inspect active work. A task ends as completed, failed, or cancelled. |
| Task queue | The work SoAI has accepted but not yet dispatched. An accepted inference request becomes a durable database record first, then waits in a bounded in-memory scheduling queue, and a full queue produces an explicit refusal rather than an unbounded backlog. See Scheduler and runtime tuning for the limits. |