Categorizers

Documentation Index

Fetch the complete documentation index at: https://docs.inboxmate.psquared.dev/llms.txt

Use this file to discover all available pages before exploring further.

Categorizers tell InboxMate how to handle each incoming message. Define a list of categories with descriptions, and the AI assigns every new email (or inactive widget chat) to exactly one. Each category can trigger actions like archiving, creating a ticket, sending a notification, replying with AI, or calling a webhook. You manage categorizers from Settings → Kategorisierer. InboxMate ships with one default email categorizer per workspace, pre-filled with seven common categories (e.g., Kundenanfrage, Intern, Spam, Newsletter, Zustellfehler).

​

How it works

1. A new email arrives, or a widget chat goes inactive (30 minutes without a new message)
2. InboxMate runs the matching categorizer through the LLM once per subject
3. The LLM picks one category, returns a confidence score, summary, and reasoning
4. Each action on the matched category runs in auto mode (executes) or suggest mode (waits for human review)

​

Categories

A category is a label plus a short description that tells the AI when to pick it. Good descriptions are one or two sentences focused on intent, not keywords:

Kundenanfrage — Eine Frage oder Beschwerde von einem (potenziellen) Kunden, die eine Antwort braucht.

The default German set covers most B2C inboxes out of the box, but you can rename, remove, or add categories at any time.

​

Actions

Each category can trigger one or more of these actions: Conflicting actions are blocked at save time — you can’t combine Archivieren with Endgültig löschen, and the same action type can only appear once per category.

​

Suggest vs auto mode

Every action has a mode toggle: automatisch tun (the system runs it) or nur vorschlagen (the action waits in the inbox for an operator to approve or dismiss). Suggest mode is the safe default while you’re still tuning a category. Mit KI antworten is the one action that runs in both modes: in suggest mode it pre-generates a draft so operators can review and send with one click; in auto mode it sends without review.

​

Confidence behavior

The behavior preset controls how confident the AI has to be before it acts: Below the match threshold, the email lands in the inbox uncategorised. Between the two thresholds, actions are suggested but not auto-executed.

​

Scope

By default a categorizer runs on every connected mailbox plus inactive widget chats. Under Gilt für, restrict it to specific mailboxes by selecting their chips, or toggle the widget-chat scope off. Workspaces with multiple mailboxes (e.g., support + sales) often want one categorizer per inbox so each gets its own categories and actions.

​

Knowledge

Two knowledge inputs are available per categorizer:

• Wissensquellen (immer mitgegeben) — up to 10 short knowledge entries injected into every prompt. Best for facts that are always relevant (return policy, business hours, escalation routing).
• Wissens-Buckets (bei Bedarf durchsucht) — vector buckets the AI can call via search_knowledge if it needs more context to decide a category. Best for large knowledge bases.

See Data sources for how to build knowledge entries and buckets.

​

Activity log

The Letzte Aktivität section lists recent categorizations and the actions taken. Reversible actions (like archive) show a Rückgängig button so operators can undo a wrong call without leaving the page.

​

Plan availability

Categorizers are available on Pro and Business plans. Auto-mode actions (including Mit KI antworten set to auto) require Business.

​

EU model routing

Workspaces with EU-only processing enabled (Pro and Business) automatically route categorization through the EU deployment of GPT-4.1 or GPT-5-mini. No configuration needed — the routing happens transparently per request.