A company doesn’t let a model answer freely about its orders. The idea is simple: spot what is asked, look in the allowed data, write one sentence, and point to a human if it goes off track.
Here a Qwen 2.5 0.5B is customised as the vildren model. It may only rephrase a fact sheet. Off topic, tampering, invention: the email, the model is not called.
It runs on a Raspberry Pi 4. The model is small on purpose: it cannot afford to make things up, so the pipeline does the work.