Settingsintermediate

Detection rule types reference

The four detector types Sidekick can use to find structured data in messages — IBAN, Phone, Domain and Custom regex. What each one matches and stores, when it fits, and how a new rule is started.

4 min read

Detection rule types reference

A detection rule tells Sidekick what structured data to look for in conversation text. When a rule matches, the value is written to the custom field the rule targets, and listed in the Detected Insights module in the Sidekick panel.

Four detector types decide what counts as a match. Pick the one that fits your data shape.

What text gets scanned

By default, detection rules scan inbound conversation text — messages from the customer. Agent replies are never scanned. You can include more conversation context with rule settings:

  • Turn on Scan internal notes to include internal notes in the scan text. Only notes written by a person are included; system, automation, Sidekick and integration notes are excluded so detection never runs on Atender’s own output
  • Turn on Scan conversation subject to include the conversation subject in the scan text

Messages and notes are read from a per-conversation high-water mark, so each run only sees what arrived since the last one. The subject has no timestamp to compare against, so it is re-scanned on every run — which means a subject-derived value is re-asserted each time and can overwrite a later manual edit of that field.

Editing only the conversation subject does not trigger a fresh detection run. The subject is included when a rule runs, but a subject-only change will not rerun detection by itself.

IBAN

International bank account numbers from any of the roughly 70 IBAN countries. Validates the country prefix, the length and the mod-97 checksum, so typos and made-up numbers are rejected. No configuration.

  • Normalized, spaces stripped — Pay to NO93 8601 1117 947 → NO9386011117947
  • — NO9386011117948 (bad checksum) → no match

Can target a Text field.

Phone

Phone numbers via libphonenumber. One setting, Detection mode:

  • Worldwide — international format only: the number must start with + and a country code. Highest precision, but won’t catch locally-formatted numbers like 22 22 22 22
  • Regional — pick one or more regions. Locally-formatted numbers from those regions are detected, plus all international-format numbers. At least one region must stay selected

Spaces, dashes and parentheses inside the number are tolerated. Values are stored in E.164 (+47 22 22 22 22 → +4722222222).

Can target a Phone field or a Text field.

Domain

Plain domain names typed into a message — acme.no, www.acme.no, example.co.uk, mail.acme.no. Validated against the Public Suffix List, so foo.fakeextension is rejected. No configuration.

Stored as the registrable domain: mail.acme.no becomes acme.no, and a leading www. is stripped. Domains inside a full URL (https://acme.no/page) or an email address (user@acme.no) are deliberately not matched.

Can target a Text field.

Custom regex

For everything else — invoice numbers, order IDs, serials. You supply a Pattern and Flags (gi by default), and the rule runs it against the configured scan text.

  • Identifiers with a consistent prefix — \bORD-\d{6,10}\b for order numbers
  • Existing regex you already maintain elsewhere — \bSO\d{6,}\b for sales orders
  • Migrating rules from another system — Direct port

Notes:

  • The rule stores the first match. If the pattern has a capture group, the first group is stored; otherwise the whole match is
  • Use word boundaries (\b) to avoid matching substrings
  • The editor carries a live tester under the Pattern field — paste sample text and matches are highlighted with a count. Use it before saving (see Create a detection rule)
  • Saving runs a safety check: an invalid pattern, or one the checker judges potentially catastrophic (ReDoS), is refused

Can target a Text field.

Starting a rule

New rule opens on What do you want to detect?, which offers three ways in:

  • describe what you want in the box and click Suggest — the AI picks the detector type and its config, and drops you into the editor with them filled in
  • pick one of the four types as a starting point
  • Skip — start blank, which opens the editor on Custom regex

Picking which type

  • Is it a bank account number? — IBAN
  • Is it a phone number? — Phone
  • Is it a bare domain name? — Domain
  • Anything else with a consistent shape (order IDs, invoice numbers, serials) — Custom regex
  • Is the data described differently every time, with no consistent shape? — None of these — detection rules are deterministic matchers, not an LLM extractor

Ordering and precedence

Rules run in list order, top to bottom, and every enabled rule that matches writes. If two rules target the same field, the one further down the list runs last and wins. Reorder by dragging a row’s handle on the Detection Rules list.

What the confidence threshold does not do

Settings → Sidekick → General carries Auto-save detected values, Show inline highlights and a Confidence Threshold slider described as gating auto-save (green) against needs-review (amber). None of the three currently affects detection rules: matches carry no confidence score, and the worker writes every match to its target field regardless of those settings. The threshold does apply to Sidekick’s AI auto-tagging, which is a different feature.

See also

Tags

Reference
Back to Settings