AI that doesn't make things up: how Welya keeps answers grounded
Language models are great at stringing words together, but a secretary who invents deadlines is a disaster. Here's how we prevent it.
· 6 min read · Welya AI team
Every request to Welya's model includes a context block built from your account: open tasks with deadlines in local time, today's schedule, free slots, unprocessed inbox items and relevant documents. The model is instructed to answer only from that block.
Replies are requested in a strict JSON schema: the answer text, a list of references (task/document ids) and actions the app can execute. Before anything is shown, the server verifies that every referenced id really exists. References to tasks that never existed are dropped.
If the model service is unavailable or the answer fails validation, Welya falls back to a deterministic rule engine that produces an equally useful reply, and labels its source. You never see an empty screen.
This approach makes Welya trustworthy for the things that matter: dates, course names, and what you actually need to do today.