Ask the assistant a question

Get a useful answer out of Ask Prodigy Genie, and know which part of what comes back is drawn from your organisation's own records and which part is the model writing prose.

Who can do this

• Front-line staff
• Site director
• Regional director
• Organisation admin
• Super admin

The screen's own check matches its rule exactly, the only screen in PTRS where that is true. A read-only account gets past the outer gate and is stopped by the screen itself with a message saying so. A guardian account is redirected away before the screen loads.

Before you start

Nothing to prepare. There is nothing to save and nothing to set up first.

Know the two limits. Twenty questions a minute across the whole module, and five answers streaming in at once, both counted per person. Past either you will be asked to wait.

Steps

1. Open the chat

Two panels: the conversation list on the left, the chat on the right. The heading reads Ask Prodigy Genie.

2. Press New Chat, or just start typing

New Chat opens an empty conversation and makes it the active one. If you type into the box with nothing active, one is opened for you.

Checkpoint

The conversation appears at the top of the left-hand list, titled New conversation. It is renamed to the first sixty characters of your first message once you send one.

3. Type the question and press Enter

The box is labelled "Ask about attendance, compliance, staffing…". Enter sends; Shift+Enter adds a line.

The answer streams in word by word. A Stop control cancels it.

4. Read the answer, then read the panel underneath it

Most answers arrive in two parts. The words are the model's. The chart, table or set of cards underneath is built afterward from a direct, live look at your organisation's own records, and the model never sees or writes it.

The panel is the part to trust for a number. Which one you get is decided by the words in your question, in English, on the server. See What the assistant can answer.

5. Use a follow-up suggestion if one appears, but repeat yourself if you build on it

Each panel carries up to three suggested follow-ups. Pressing one sends it as a brand new, independent question with no memory of the one before it.

Checkpoint

You have an answer, and for most kinds of question, a panel under it. Nothing has been saved anywhere.

Result

An answer on screen. That is the whole of it. See what this does not do, below.

Phrasing that reaches the right data

The server decides which of 23 kinds of question you asked before it fetches anything, and that decision is what decides what data ends up in front of the model. Two questions that mean the same thing to a person can be sorted differently.

To get	Include a word like
Live headcount and per-site capacity	attendance, headcount, checked in, who's here, ratio
Incident counts by type and severity, and the OCCL backlog	incident report, classify incident, draft report
A DELACARE or OCCL answer from loaded documents	regulation, DELACARE, OCCL, rule, requirement, section
Compliance scores per site and open alerts	compliance report, compliance summary, inspection
The organisation's headline numbers	board report, executive summary, weekly summary
Incident trends and site hotspots	pattern, trend, anomaly

Two things follow from the way this works, and both are worth knowing.

An unrecognised question is answered as though it were about attendance, because that is the fallback used whenever nothing else matches.

Which panel you get is decided by matching pieces of your actual wording, so a word can match by accident. Asking for "the accurate headcount" can return the check-in-rate chart, because the panel picker is matching the letters inside accurate as well as the word rate on its own.

What this does not do

Nothing about this conversation is kept

The conversation list, the messages and the titles live only in your browser. Refresh the page and everything is gone.

The assistant cannot follow up. The question you send carries only the text you just typed, with nothing about what came before it. "Show me more" is answered as though it were the first thing you had said. Repeat the subject in every question.

Nothing is recorded on PTRS's side either, so there is no history for anyone to review later, yours or an administrator's.

Three answers you should not act on without checking

Any attendance figure stated in words. Almost every one the assistant can state comes from a summary table nothing currently fills in, so it is zero. The one exception is the live headcount panel, which reads today's check-ins directly. See Attendance.

Any regulation quotation. The retrieval step knows which document and which section each passage came from and the answer you see does not carry it. And if nothing has been loaded, the model is told to answer from its own general knowledge; nothing checks that it actually says so.

Any percentage that names one site. Five panels answer an organisation-wide question about the first site alphabetically: the 14-day trend chart, absent members, check-in rate, room breakdown and the streak leaderboard. Each is labelled with that site's name; the words above it cover every site.

Common errors

What you see	Why	What to do
A message saying you do not have permission to access the AI Assistant	You hold a read-only account. The outer gate lets that role reach the address; the screen's own check refuses it, correctly	Nothing here. A read-only account satisfies no rule anywhere in PTRS
A message saying the AI service is unavailable, and nothing streams	Either the model cannot be reached, or the outside service is selected with a blank access key, in which case every attempt fails and is retried twice before giving up	Open the model status screen if you can, or ask an administrator to. The provider name and the two status dots are the whole diagnosis
The answer stops mid-sentence and resumes oddly	Each streamed piece of the answer is sent with no allowance for a line break inside it, so a line break inside one piece cuts off the rest of it. Headings and lists are the usual casualties	Ask for the answer as a single paragraph
An email, a phone number or a date comes back replaced with a redaction marker	On the streaming path, real values are restored before the answer is checked rather than after, so a value that was correctly restored is then caught by that same check. The date check matches any date written in year-month-day order	Expected. See How PII is protected
A message asking you to slow down	Twenty questions a minute, or five answers streaming at once	Wait
A suggested question about translating a message returns an answer about headcounts	That suggestion's own purpose-built action is not reachable from any screen. Typed into the chat, it is answered as a translation question with three general, organisation-wide counts instead of the message you meant to translate	Paste the text you want translated into the same message

Related

• What the assistant can answer, all 23 kinds of question, the data each one sees, and the panels it can produce
• How PII is protected in the AI pipeline
• Who can do what with the AI Assistant
• AI troubleshooting