MATE AI
The assistant runs in a right-hand panel: context-aware, restricted by a data wall, and configured per account.
| Element | Behaviour |
|---|---|
| Context | Attached automatically from the URL: current page, log, and the module open on the page. |
| Starters | Suggested questions that change with context; the strip hides itself once you have asked there. |
| Answers | Streamed over Server-Sent Events, with navigation chips that deep-link into the app. |
| Guidance cards | Per-module and per-process explanations you can generate, refresh, or regenerate; cached with the module's results. |
| No provider | The panel says so and links to Settings → AI; the starter chips are disabled. |
Providers#
| Provider | Notes |
|---|---|
| Anthropic | Native transport; no embeddings. |
| OpenAI | Also used for embeddings. |
| UniGPT | Institutional gateway; base URL required. |
| Custom | Any OpenAI-compatible endpoint. |
The form takes a key (masked afterwards; leaving it blank keeps the stored key), a base URL where needed, then Fetch models to populate the pickers. When an administrator controls the AI configuration, the panel is read-only and says so. Modules can go further: ai_models in their manifest renders their own provider and model selectors, and self_hosted: true gives a module an isolated key field.
The data wall#
One switch decides how much context is used, off by default: access process data enables both the process list (which logs exist and their headline figures) and activity and variant aggregates, capped at 40 activities and the top 15 variants and skipped for OCEL logs.