Learning your product (domain learning)

Give TR-AI your product URL and a login (use a dedicated, read-only demo account). TR-AI opens the product in a headless browser, logs in through your normal login form and explores breadth-first: navigation menus and dropdowns, tabs, accordions, settings sub-pages, “Load more” lists and read-only “View / Details” dialogs — up to 1,000 screens within the time budget (60 minutes by default). Long lists (tickets, invoices, projects) are sampled — up to 25 screens of each kind — so the budget covers every area of the product instead of one list.

What TR-AI learns

• A product map: every screen type, grouped by product area, with the menu path that reaches it (for example Settings › Billing › Invoices), built from breadcrumbs and the menus and links it followed.
• A screen card for each screen type: headings, form fields (label, type, required, help text, the options of each drop-down), button labels, table columns, status badges, tooltips and empty-state messages. Buttons are read, never pressed.
• Workflows: step-by-step tasks such as “How to add a user”, written by AI from the screen cards, the menu paths and your knowledge-base articles. Each step uses only menus, fields and buttons TR-AI actually saw.
• A glossary of product terms and the other words customers use for them.

When a customer asks how to do something, TR-AI prefers a matching workflow and answers with the exact menu path and field and button names. It is told never to mention a screen, menu or button that is not in what it learned.

Review workflows

1. 1
   Open Domain learning

   The page shows the crawl phase while it runs (logging in, exploring, understanding screens, writing workflows, embedding), then the last crawl’s coverage: screen types, product areas, workflows and AI calls.
2. 2
   Check the product map

   Open an area to see its screens and their menu paths. A missing area usually means the demo account cannot see it.
3. 3
   Edit or disable workflows

   Edit a workflow’s title or steps (one per line) to make it human-approved — re-trains never overwrite it. Disable one to keep it out of answers. Changes apply to answers immediately.

• JavaScript apps are fully supported: each screen is rendered before it is read.
• It follows links across your product’s own sub-domains (e.g. portal.acme.com → app.acme.com), never other websites or your marketing site.
• It never presses buttons that change data (save, delete, create, invite …) and never opens links like log out, delete, edit, archive or unsubscribe — but use a demo account anyway.
• Personal data shown on screens (e-mail addresses, phone numbers and similar) is masked before anything is stored or sent to the AI. Screen cards and workflow steps are encrypted at rest.
• Re-trains are incremental: only product areas with new or changed screens get new workflows, which keeps re-train AI cost low (at most 60 AI calls).
• If the session expires it logs in again; SSO and two-factor logins are not supported (use a password-only demo user).
• When only a few screens are found, the Domain learning page explains why (login rejected, JavaScript screens without the browser, errors, links to other sites).