What an iFood user representation looks like
An LCM user representation is a structured markdown document with about ten sections, generated by a single LLM call from one user’s data. Below are three real excerpts from a generated profile for one iFood user, so you can see what the output of all this looks like before you start preparing your own input.
User Profile & Persona
A few sentences that fingerprint who the user is — their motivation, their decisiveness, and what kind of orderer they are. Downstream tasks use this to set tone, segment, and message style.
1) User Profile & Persona
User Persona
Targeted, high-basket orderer for group dinners. Search-driven for specific cravings, loyal to proven merchants, optimizes for meal voucher acceptance and delivery fees. Browses extensively in categories they never purchase.
Motivation
Feeding multiple people for dinner. Most orders contain multi-person portions (pizza combos serving 4–6, 30+ esfihas, multiple burger combos, paired ramen bowls).
Decisiveness
Specific searches (“Imigrantes”, “Poke”) convert quickly. Category browsing leads to extensive comparison — sometimes across weeks — before purchasing or abandoning.
Food Preferences, Categories & Top Merchants
Where most of the personalization fuel lives. Top dishes at the category level, browse-vs-buy gaps, ingredient preferences. Re-ranking, query expansion, push targeting, and filter suggestions all draw from this section.
4) Food Preferences, Categories & Top Merchants
Top Categories (dish-level)
- Pizza — large family-size, prefers thin crust, recurring Friday dinner choice
- Lanches — smash burgers and double cheeseburgers from Dingo Smash Burger, Jungle Burguer
- Arab cuisine — bulk esfiha orders from Habib’s Esfiha — Augusta, paired with quibe
Browsed but never purchased
Repeatedly views Japanese (poke, temaki) and healthy bowl options — consistent comparison shopping without conversion. Most likely a partner’s preference not aligned with group-dinner format.
Favored ingredients
Avoided ingredients
Segmentation Signals
A structurally different section at the end of the profile. Three dense, keyword-rich plain-text blocks — no formatting — designed for vector embedding and ad-hoc audience matching. Same signal lives elsewhere in the profile; this is an experiment in concentrating it for downstream retrieval.