Fixed replies for red-flag phrases

Some messages need an answer nobody improvises. When a patient writes “my mum’s face is weak on one side”, a clinic wants its own approved words sent back (“go to the Emergency Department now”) and a person alerted, every time. A phrase rule does that: a message containing one of the rule’s phrases gets the rule’s reply exactly as written, the AI is not asked, and the rule’s Live signal is raised.

You need: an agent with a Live signal for it (for example “Clinical red flag”, see Choose what calls find out), and an owner or admin role.

What a rule matches, and what it misses

A rule matches words, not meaning. Before comparing, case, punctuation, emoji and Latin accents are ignored, so “FACE IS WEAK ON ONE SIDE!!” matches “face is weak on one side”. It still misses:

• misspellings (“face is waek”)
• other ways of saying it (“my mother’s face is drooping”)
• a phrase split across two messages

A missed message goes to the AI as usual. Keep the Live signal on as the safety net, and list the few ways people really say it, in each language your customers write in. Every language list is checked against every message: the language of a short message is not guessed.

Three match modes:

mode	matches
word (default)	the phrase’s words in a row, as whole words: “chest pain” matches “bad chest pain”, not “chest painting”
contains	anywhere, even inside a word
exact	the whole message, nothing else

A Chinese, Japanese or Thai phrase is matched anywhere in word mode, because those languages have no spaces between words.

Exclusions veto a rule for a message: with the exclusion “no chest pain”, “I have no chest pain today” does not match “chest pain”.

Set one up in the dashboard

1. Open the agent, then Channels → Fixed replies for set phrases.
2. Add rule: a name, the phrases per language, the approved reply (and the reply in each other language you listed), the Live signal, and whether a match hands the conversation to a person.
3. Turn on Red flag for a clinical or safety rule.
4. Type a few real messages into Test a message: it shows which rule matches and the reply that would be sent. Nothing is sent or recorded.

Write every reply yourself and have a person approve each language. The reply is sent as written; it is never translated or reworded.

What happens on a match

• The agent is answering: the reply goes out, the Live signal request is recorded (detected_by phrase_rule) and your alert workflows fire. The AI answers the next message as usual, unless the rule hands off.
• A person has taken the conversation over (or the AI is paused, or a workflow owns it): nothing is sent. The signal is still raised and the conversation timeline shows the rule that matched.
• Several rules match: each raises its signal; one reply is sent, a red flag’s first, then the lowest priority.

A red-flag rule must raise a Live signal. Its alert is never held back by the limit of five alerts per conversation a day. If the signal cannot be raised (it was switched off, or the workspace hit its hourly alert ceiling), the conversation is handed to a person instead, which alerts the team.

Phrase rules work on WhatsApp, SMS, Messenger, Instagram, Telegram, LINE and the website chat. Not on email or calls yet. In the dashboard’s test chat a rule answers but nothing is recorded and nobody is alerted.

With the API

curl -s -X POST https://api.omniocto.com/v1/agents/ag-0000000001/phrase-rules \
  -H "Authorization: Bearer $OMNI_KEY" -H "Content-Type: application/json" \
  -d '{
    "name": "Stroke signs",
    "phrases": [
      {"lang": "en", "text": "face is weak on one side"},
      {"lang": "en", "text": "face weak on one side"},
      {"lang": "zh", "text": "半边脸没力"}
    ],
    "reply_text": "This may be a stroke. Please go to the nearest Emergency Department now, or call 995.",
    "replies": {"zh": "这可能是中风。请立即前往最近的急诊部，或拨打995。"},
    "signal_field_key": "clinical_red_flag",
    "red_flag": true
  }'

GET /v1/agents/{id}/phrase-rules lists them. PATCH /v1/agents/{id}/phrase-rules/{rule_id} changes one (send the version you read as expected_version; enabled: false switches it off), and DELETE removes it. Writes need an owner or admin key. Every field: Lookups reference.