The objects that make up the platform and how they relate.
A workspace belongs to one company. It holds users, roles, knowledge bases, agents, API keys and billing. Enterprise customers can have several workspaces (for example per subsidiary) under one organisation.
| Role | Can |
|---|---|
| Member | Ask questions in channels and the web app; see their own conversations; trigger agents they've been granted |
| Editor | Everything a member can, plus create and edit agents and knowledge bases they own |
| Admin | Manage users, roles, sources, apps, API keys and billing; see all run traces and audit logs |
| Owner | Admin plus contract, data deletion and workspace transfer |
A named, permission-aware index of content from one or more sources. Access to a knowledge base is granted to users, channels and agents. A question is answered from the knowledge bases the asker can access, filtered further by the asker's permissions in each source.
A connection to a system (Google Drive folder, SharePoint site, Notion workspace, Zendesk account…). A source belongs to exactly one knowledge base and syncs content and permissions on a schedule or via change notifications.
A document is one item from a source (a file, a page, a ticket, a Slack thread). Documents are split into chunks for retrieval. Citations point to a chunk and, where the source supports it, to the exact location inside the document.
One question to the Knowledge Search Agent, from any channel or the API. A query produces an answer, citations and a confidence score, and is logged.
A thread of queries and answers that share context — a Slack thread, a Teams chat, a web-app conversation, or an API conversation you create.
A configured worker with a job (instructions), knowledge bases it may read, tools with scopes, guardrails, triggers and optionally an evaluation set. Agents are versioned; a version is promoted from test to production.
A capability an agent can use: read a mailbox, write a row, post a purchase invoice, call an HTTP endpoint. Tools come from integrations and have separate read and write scopes.
One execution of an agent, from trigger to completion. A run has a trace — every retrieval, tool call, decision and approval — and a status: queued, running, waiting_approval, succeeded, handed_off, failed, cancelled.
A pause in a run where a policy requires a person to allow a tool call. Approvals are answered in Slack, Teams, the web app or via API.
What starts a run: a schedule, an event from a connected system, a webhook, a mention in chat, or a manual start.
An immutable record of security-relevant events in the workspace: sign-ins, permission changes, source connections, API key actions, agent promotions, emergency stops. Exportable to your SIEM.