Prepare your data for your first user representation
An LCM user representation is one user’s commerce life written as text the model can read: who they are, then everything they did over time, in order. Your only job before the bootcamp is to get that text ready for at least one real user. We turn it into a representation together when you arrive.
See an example of a user representation Real generated profile excerpts — persona, food preferences, segmentation block. See what we use representations for Three downstream surfaces with real profile excerpts — re-ranking, push, query expansion.- Pick one real usersomeone with a decent history
- Gather their historyorders, then searches and browses, in time order
- Add a short glossary and save itone text file, ready to bring
Pick one real user
Choose one real user with a decent history, ideally over the last year or so. The data of a single user is enough to start ideating on representations. A heavy user with lots of activity makes it easier, since the data is richer. You can also pick yourself, using your own data from your company.
Gather their history
Pull what you already have for that user and lay it out in time order. Start with orders. If it is easy, add the things they did around buying: searches, browses, items viewed, abandoned carts. Keep it readable. Do not spend time engineering a format.
2024-09-06 Fri Dinner Habib's Esfiha Esfiha Carne x30 R$ 245.40 PIX (first order here)
Date, when, where, what, how much, how paid, and a short note if something stands out. That is plenty.
Add a short glossary and save it
At the top, write a few lines explaining the terms an outsider would misread: payment types, acronyms, subscription tiers, shift or time definitions, status codes. Keep it short, and only cover what could genuinely be misunderstood.
Clube paid subscription with recurring discounts; a committed user VOUCHER employer meal benefit used as payment, not a discount Dinner order shift; Friday Dinner counts as weekend (first) first order from this merchant
Put the glossary on top, the user’s history below, and save it as one text file. That file is what you bring.
- A short glossary of your domain terms
- That user’s history in time order: orders, plus searches and browses if you have them
GLOSSARY Clube paid subscription, recurring discounts (a committed user) VOUCHER employer meal benefit used as payment, not a discount Dinner order shift; Friday Dinner counts as weekend USER 48213 - Sao Paulo - Clube member - ~14 months of history 2024-09-06 Fri Dinner Habib's Esfiha Esfiha Carne x30 R$ 245.40 PIX 2024-09-13 Fri Dinner Madero Classic Burger R$ 62.00 CREDIT 2024-09-20 Sat searched "sushi" -> no order 2024-10-02 Wed Lunch Acai da Barra Bowl 500ml R$ 24.00 VOUCHER (first order here) ...
A few lines of glossary on top, then that user’s history in time order. That is the whole thing. The exact format does not matter yet; we shape it together at the bootcamp.
If you have the time, you can start expanding on what you just built. Expand the input data, following the iFood example we provided — add aggregated user statistics and other signals beyond purchases. You can also expand the sample to more than one user.