mirror of
https://github.com/mauriceboe/TREK.git
synced 2026-10-18 21:16:47 +00:00
Maurice
41b14ebc2a
Gemini behind the OpenAI-compatible provider decodes against the JSON schema TREK sends and fills declared properties only. reservationFor was an open object without any, so the model left it out: a Booking.com stay came back with its dates and price but no hotel name, and the mapper had to drop it (#2477). It is the same cause as #1638, whose fix only made the JSON readable. The schema now declares every field the prompt names inside reservationFor, for venues, flights, trains, buses, boats, events and rental cars, at most two levels deep, and requires it; the car desks at the root get the same venue fields. Every object stays open to keys the prompt does not name. On the cloud AI path an answer is cleaned up before it is mapped. Exact duplicate nodes are dropped, so one stay is one booking. priceCurrency is read as an ISO 4217 code ("EURials" is EUR, a lone symbol maps to its code); when the model names none, a priced node takes the currency of the document's total, and otherwise the field goes instead of saving a code nobody can convert. The kitinerary path is untouched. The import upload decodes file names as UTF-8 like every other upload route. "Bestätigung.pdf" reached the warnings as mojibake, and the review could not attach the file to the booking it created, because it matches on the name. The OpenAI-compatible client logs which response_format the answering attempt carried, metadata only, so the next report shows whether the schema reached the model at all. The background import card showed a green tick beside "no reservations could be extracted". A parse that found nothing now wears the warning mark instead; it is the same card on the desktop and on the phone.