Skip to content

AI personalisation

Cue can rewrite a message's title and body for each recipient — warmer, in their language, mentioning their name — using any model supported by Pydantic AI: OpenAI, Anthropic, Google, Groq, Mistral, Ollama and OpenAI-compatible gateways, among others.

It is off by default and designed to be safe for transactional and financial messaging.

Enable it

pip install 'cue-notify[ai]'

[ai]
enabled = true
model = "openai:gpt-5-mini"                   # or "anthropic:claude-haiku-4-5",
fallback_models = ["google:gemini-2.5-flash"]  # "ollama:llama3.2", …
instructions = "Friendly and concise. Never use exclamation marks in SMS."
daily_token_budget = 2000000

Credentials are read from each provider's standard environment variable (OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, OLLAMA_BASE_URL…).

Then opt templates in:

{
  "key": "renewal-reminder",
  "locales": {"en": {"title": "Your plan renews soon", "body": "Pro renews on {{ data.date }} for {{ data.price }} USD."}},
  "ai_instructions": "Mention how long they have been a customer if known.",
  "ai_attributes": ["first_name", "customer_since"]
}

Guardrails

  • The template always renders first. The model receives that text and may only rephrase it; links, images and data payloads never come from the model.
  • Facts must survive. Every number and URL of the original must appear in the rewrite, and no new ones may be introduced (1 250 000 and 1,250,000 are treated as equal). Violations are sent back to the model to fix; if it still fails, the template text is used.
  • Length limits relative to the original keep SMS short.
  • Minimal data sharing. Only the attributes in ai_attributes are sent, never the full profile, addresses or event payload.
  • Prompt-injection resistant framing. Original text and recipient context are passed as data, and the model is instructed to ignore instructions inside them.
  • Never blocks delivery. Timeouts, provider errors, budget exhaustion — the message is sent with the template text.

Personalisation happens at send time (after quiet hours and caps), so tokens are never spent on messages that end up suppressed. Messages record generated_by (template or ai:<model>).