Moments AI module

The Moments AI module adds on-device AI inference to mobile apps using the Tealium Prism SDK. It runs locally on the device using Gemini Nano with no server-side component. Configure the module through the Mobile Configuration slideout in Tealium iQ. The module is available only in mobile profiles.

Module configuration

When adding the module, configure the following settings:

• System Prompt: Define the AI model’s role and output format. This prompt sets the overall behaviour and constraints for every inference. Example: instruct the model to return JSON in a specific schema, or to act as a particular type of assistant. Supports variable substitution using {{handlebars}} syntax.
• User Prompt: Provide the per-request instructions sent to the model. Use {{event.}} to insert values from the current event (e.g. {{event.user_name}}), and {{history}} to include recent event history. These placeholders are replaced at runtime with real data from the mapped context. Supports variable substitution using {{handlebars}} syntax.
• Temperature: Numeric input for model temperature.
• Max Tokens: Max output tokens for the on-device model.
• Max Event History: Number of recent events to retain as context for the model.

Variable substitution

The System Prompt and User Prompt fields support handlebars-style {{variable}} substitution at runtime. The module resolves these placeholders before sending the prompt to the on-device model.

The following built-in variables are available:

Custom variables are also available through prompt variable mappings.

Mappings

Configure mappings in the Mappings section of the Moments AI module in Tealium iQ.

Command triggers

Map event conditions to SDK commands. When the mapped condition is met, the specified command runs. To map events, refer to Create an Event Mapping

Config overrides

Map data layer values to override module configuration fields at runtime. The following destinations are supported:

Context mappings

Map variables to capture into the stored event history. Enter the context key name to store in event history. For example, mapping to checkout produces context.checkout.

Prompt variable mappings

Map data layer values to custom variable names for use in prompt substitution. After mapping, reference the variable as {{variable_name}} in the System Prompt or User Prompt fields.

Example

The following example uses Moments AI to generate a personalized cart abandonment notification.

Configuration:

System Prompt:

You are an on-device e-commerce personalization assistant for a mobile retail app. Your job is to turn a shopper's current activity and recent behavior into a single, timely next-best-action.
Respond with ONLY a compact JSON object, no prose, no markdown, matching exactly:
  {"headline": string, "message": string, "recommended_category": string, "urgency": "low"|"medium"|"high"}
Rules:
  - "headline": <=6 words, action-oriented.
  - "message": <=160 characters, warm and specific to what they browsed or bought. Never invent products, prices, or discounts that aren't in the provided data.
  - "recommended_category": choose from the categories the shopper has shown interest in; if unknown, use "trending".
  - Do not repeat back email, user IDs, or any personal identifiers.
  - If you have insufficient signal, set urgency to "low" and keep the message generic but on-brand.

User Prompt:

Current event: {{event.tealium_event}}
Shopper location: {{event.customer_city}}
Products in view/cart: {{event.product_names}}
Categories: {{event.product_categories}}
Brands: {{event.product_brands}}
Price(s): {{event.prices}} {{event.currency}}
Recent activity (most recent last): {{history}}

• Temperature: 0.7
• Max Tokens: 150
• Max Event History: 5

Mappings:

• Command triggers:
  • event_name == "product_view" → store_event
  • event_name == "app_background" → trigger_inference
• Context mappings:
  • product_name, product_price, and page_name → context
• Prompt variable mappings:
  • cart_contents → cart_contents

When the app_background event fires, the module sends the configured prompts to the on-device model. It injects the last five stored events as {{history}} and the current cart contents from the data layer as {{event.cart_contents}}. The model returns a JSON response the app parses to display a personalized push notification.