An order that lives in a table between runs
Every message about the order appends a row, and the bot answers knowing its whole history.
What the scenario looks like
Сообщение о заказе
Что произошло▶
Модель▦
Что было раньше▶
Новая запись▶
Запись события▶
Ответ по заказу▶▦
Ответ▶
- Execute + Data
- LLM
- Execute
Every message is a separate run, and runs know nothing about each other. An order’s history exists precisely because somebody records it: with no collection, “what’s happening with my order” can only be answered with “tell me what happened so far”.
What is inside
- A message about an order — the entry.
- What happened — the model pulls the number, the new status and the gist.
- A new record and Store the event append a row.
- What came before reads everything already recorded for that order.
- Answer about the order confirms the change and recaps the order’s path by date.
Append, do not overwrite. Storing a single “current status” is cheaper, but then “when was it shipped” has no answer.
What to set up for yourself
- The list of statuses in the field description: it maps any phrasing onto your vocabulary.
- Filtering by order. It currently reads the last fifty events across all orders and lets the model sort it out. Once there are many orders, add a condition on the number — otherwise it is both expensive and imprecise.
- Who sends the events. In practice not a person but a webhook from your own system; the graph does not change.
What to connect
A collection under Tables. Nothing external: the whole history lives inside the product.