AI, MCP and knowledge

Diting's AI is not a separate chat window. It runs on top of product pages, reads their state and executes registered actions.

Page contracts

Every product page registers two things:

• State AI can read: current filters, time range, selected objects
• Actions AI can run: open, filter, drill down, change range

Pages without an adapter expose only title and path.

Permissions

AI reuses product RBAC. Whether an action runs is decided by role; anything beyond it is rejected and logged.

Actions

• Only registered actions are available
• Each execution carries a state version, so a changed page cannot be acted on by mistake
• Results are written back to the page and into audit

MCP and knowledge

• MCP servers let AI call external tools
• The knowledge base answers product and operations questions
• Both are bound by the same permissions

Sandbox

Code and commands run in a sandbox, and /workspace persists. The sandbox is not VM-grade isolation; do not run untrusted binaries in it.

Models

• Multiple providers
• Local models
• Offline model import for isolated sites

Audit

Actor, session, action, parameters and result are recorded and searchable.

Related: Access, audit and tenancy, API and OTLP.