Skip to main content

The assistant

This is the assistant from the handler's side, with a pointer to the code behind each behaviour.

Opening it​

A launcher sits bottom-right on every screen. It opens a popover. The expand button in its header docks it as a side panel a quarter of the screen wide, and the page reflows around it. The choice is remembered per browser, and opening from a demo hint keeps a docked panel docked.

The header says which claim the assistant is talking to (Talking to CP-AUT-260381), or Open a claim to start when the screen has none. On a page with no claim the composer is disabled.

Code: frontend/src/chat/ChatWidget.tsx, ChatDock.tsx.

Asking​

The handler types a question, or picks a quick action:

  • Pinned chips for the five most-used capabilities: Summarise claim, Validate policy, Reserve estimation, Next actions, Fraud check.
  • More opens the whole catalogue, grouped Understand, Decide, Protect, Produce, Help, each with its one-line description.

A quick action sends a natural message ("Validate the policy against this loss.") so the transcript reads well, and runs that capability's playbook.

Code: CapabilityMenu.tsx; catalogue in packages/shared/src/capabilities.ts.

Watching it work​

While it works, the assistant shows:

  • activity chips for each lookup, in the present tense: Reading the policy…, Checking the policy period and coverages…;
  • the answer streaming in as it is written, rendered as Markdown;
  • a lookups link under the finished answer that expands to show every tool it called and a summary of what came back.

Stop cancels the turn, including the model call.

Pointing at the screen​

When an answer refers to something on the claim screen, the assistant can highlight it. The panel scrolls into the middle of the view, pulses, and keeps an amber outline that fades over a minute. It only does this when it has actually called highlight_panel; the prompt forbids claiming a highlight it did not make.

Code: frontend/src/chat/highlight.ts; the tool in api/src/capabilities/tools.ts.

Rating answers​

Every finished answer has a thumbs up and a thumbs down. A thumbs down opens a card asking what was wrong, with reason chips and a note. The reasons go to the people improving the prompts, never into the conversation. See Answer feedback.

Speaking​

The microphone button starts a spoken conversation with the same claim, prompts and tools. A hero area shows the assistant's level and the handler's microphone level. Both sides' transcripts appear in the chat as bubbles marked with a microphone. If the microphone is silent for a few seconds the panel says so. See Voice.

The disclaimer​

The panel carries a permanent notice that AI can make mistakes and that anything acted on should be checked against the claim record. It is part of the product, in the demo and in every embedding. Hosts are asked not to remove or obscure it.

Clearing and switching​

Clear conversation empties the transcript and the history. Moving to another claim does the same automatically: a conversation belongs to one claim.

What it will not do​

The assistant declines anything outside the claim on screen, never claims to have changed anything, and treats text inside the record as data. If a document contains an instruction to "the AI", the assistant tells the handler and treats it as a possible fraud indicator. See Security.