Examples of input data based on iFood representations

A look at the actual data we feed the model to generate a user representation — click anything to see its shape.

The input to a user representation has two main pieces: aggregated stats we pre-compute in SQL (the model is bad at math — pre-compute what you can), and input types we layer end-of-funnel first (orders → carts → viewed items → merchants → searches → clicks). Below is what each looks like in practice for one iFood user.

Part 01

Aggregated stats — eight columns per user

Numeric summaries computed in SQL. Each user gets one row of these stats, grouped into logical columns. They double as prompt context for the model and as queryable SQL features for downstream filters.

user_identity_summary

Account, platform, devices, analysis window

user_location_summary

Delivery districts, merchant districts, monthly mix

user_discovery_summary

Entry points, filter usage, per-page breakdown

temporal_behavior_summary

Shift × day-type matrix

monthly_trends_summary

Per-month orders, spend, AOV, sessions

order_economics_summary

AOV, payment splits, vouchers, delivery fees

order_status_summary

Concluded / cancelled / declined distribution

browsing_preferences_summary

Top categories, recurring merchant interest

{
  "account_id": "...",
  "registration_date": "2020-09-21",
  "platform": "IOS",
  "analysis_period": {"days": 367},
  "unique_device_count": 2,
  "concurrent_device_usage": true,
  "devices": [
    {"device_name": "Apple iPhone 14 Pro", "total_orders": 72, "total_sessions": 274}
  ]
}

{
  "primary": {"city": "SAO JOSE", "state": "SC"},
  "delivery_districts": [{"district": "Kobrasol", "district_count": 155, "district_percentage": 35.07}],
  "merchant_districts": [{"merchant_district": "Centro", "merchant_district_percentage": 10.63}]
}

{
  "entry_points_for_converted_sessions": [{"entry_point": "search", "count": 84}],
  "filters_usage": {
    "summary": {"sessions_with_filters_viewed": 179, "total_filters_clicked": 5},
    "by_page": [{"page": "category_home", "top_filters_clicked": ["Ordenar"]}]
  }
}

{
  "shift_day_matrix": {
    "orders":        [{"session_shift": "Dinner", "weekday": 84, "weekend": 38}],
    "conversion_pct": [{"session_shift": "Lunch",  "weekday": 27.42, "weekend": 25.71}],
    "avg_spend":     [{"session_shift": "Dinner", "weekday": 100.20, "weekend": 115.07}]
  },
  "day_type_totals": {
    "weekday": {"orders": 102, "total_spend": 9589.81},
    "weekend": {"orders": 48,  "total_spend": 4876.51}
  }
}

[
  {"month": "2025-08", "orders": 11, "total_spend": 755.75, "avg_order_value": 68.7, "sessions": 31, "orders_weekday": 7, "orders_weekend": 4},
  {"month": "2025-09", "orders": 14, "total_spend": 1320.40, "avg_order_value": 94.3, "sessions": 42, "orders_weekday": 9, "orders_weekend": 5}
]

{
  "summary": {"window_orders": 157, "avg_order_value": 95.72, "median_order_value": 80, "avg_weekly_orders": 3.02},
  "payment_methods": [{"method": "CREDIT", "brand": ["MASTERCARD"], "pct": 45.86}],
  "vouchers": {"usage_rate_pct": 0.70, "types_used": [{"type": "clube", "pct": 90}]},
  "delivery": {"fees": {"avg_gross": 5.64, "avg_paid": 4.24}, "time_minutes": {"avg": 42.95}}
}

{
  "distribution": [
    {"current_status": "CONCLUDED", "count": 153, "pct": 97.45},
    {"current_status": "CANCELLED", "count": 3,   "pct": 1.91}
  ]
}

{
  "categories": [{"category": "Lanches", "orders": 35, "sessions_viewed": 56, "total_views": 136}],
  "recurring_merchant_interest": [
    {"merchant_name": "Dingo Smash Burger", "total_views": 20, "converted": true},
    {"merchant_name": "Jungle Burguer",    "total_views": 5,  "converted": false}
  ]
}

Click any column above to see its shape

Part 02

Input types — layered end-of-funnel first

Beyond the aggregated stats, the model gets event-level text. We layer signals from closest-to-purchase down to broadest browsing — each layer drops whatever the layer above already captured, so the model never sees the same behaviour twice. Click any tag to see what that layer looks like in the prompt.

<orders>Completed and attempted orders — the strongest signalwhat they bought

<cart_info>Abandoned carts — funnel stage, time-to-cart, error flagswhat they almost bought

<items>Items viewed but not carted — browse intent vs buy gapwhat tempted them

<merchants>Merchants visited where no item action followed — comparison shoppingwho they considered

<searches>Queries that didn’t convert — specific intent without a buywhat they were chasing

<clicks>Navigation journey — screens accessed per sessionhow they moved

Example — one order line in the prompt2025-08-15 | Fri | Dinner [TOP] [NEW]
Habib's Esfiha - Augusta 4.6 (Internal: 4.8)
  - [30x] Esfiha Carne: pão tradicional (R$ 5.50)
  - [4x] Pizza Margherita: queijo mussarela (R$ 38.00) [molho extra]
Delivery: 45mins | 2.3km | Fee: 8.99→0.00
Entry Point: home_carousel
[VOUCHER UNUSED] clube_5_off | CREDIT (MASTERCARD)
Review: "comida deliciosa, entrega rápida"
Order Rating: 5/5 | Delivery Rating: 5/5
Total: R$ 245.40 | Discount: R$ 12.00 (4.9%)

Example — abandoned cart2025-07-22 | Sun | Dinner [ABANDONED]
Pizza Hut [Initiated][Viewed][Dropped]
  - [1x] Pizza Família (8 fatias): mussarela + calabresa (R$ 89.90)
Time-to-cart: 4m20s | ETA: 50mins | Fee: 12.99
[ERR:PAYMENT]

Example — items viewed but never cartedAçaí Tradicional 500ml — Açaí da Vila (R$ 19.90)  [viewed 3x, 2 sessions]
Poke Bowl Salmão — Poke Box (R$ 42.00)         [viewed 5x, 3 sessions]
Temaki Atum — Yamato Sushi (R$ 28.00)            [viewed 2x, 2 sessions]

Example — merchants visited, no item action followedFit de Fato       [visited 3x, no item action]
Salada Express    [visited 2x, no item action]
Sucos da Praça    [visited 4x across 2 weeks, no item action]

Example — queries that did not convert"poke"           [term, global]          → no conversion
"japonesa"       [historic, global]      → no conversion
"saudável"       [trend, global]         → no conversion
"sushi delivery" [term, global]          → no conversion

Example — per-session navigation journeySession 1: Home → Restaurant Home → Search → Merchant Page  → no_conversion
Session 2: Push deeplink → Merchant Page → Cart → Checkout  → conversion
Session 3: Home → Promo Page → Merchant → Item Detail        → no_conversion

Click any input type above to see an example

Plus a short header block at the top. Before any of this, a small glossary defines the domain-specific terms the model needs to read your data correctly — Clube, VOUCHER, Dinner, anything that could be misread. Small, surgical, reused for every artifact. You’ll write yours during pre-work.