AI Assistant

Tom Bradley runs the Claymont club. Between opening the building, covering a sick call, and getting ready for an OCCL visit, he does not have time to open three different screens to answer one question. So he types the question instead: how many kids checked in this morning, whether anyone's certification is about to lapse, what DELACARE actually says about outdoor play in the heat. This page is for Tom, for any staff member who reaches for the same box, and for the one person in the organisation who loads the regulations it can answer from.

Two things to know before you rely on this for anything that matters

Read the wording on the model status screen before you type a real child's name into anything. One setting decides whether a question stays on your own organisation's equipment or leaves it for an outside service. The status screen, described below, shows which one is active right now.

The plain-English data question screen can show you another organisation's children. On a copy of PTRS shared by more than one organisation, this one screen does not stay inside the walls the rest of the product keeps around your own records. See who can do what with the AI Assistant.

What the AI Assistant is for

A question in plain English, answered from PTRS's own data where it can be, and from a language model's general knowledge where it can't. Three things it tries to do:

• Answer a question about your own organisation, on attendance, incidents, compliance or staffing, in words, with a chart or table under it.
• Answer a question about the regulations, from documents your organisation has loaded in, rather than from what the model already knows.
• Let somebody who does not know how to write a database question type one in English instead, and see what it actually asked for.

Who uses it

Who can do this

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

The chat screen's own check matches its rule exactly, the only screen in PTRS where that is true. A read-only account can reach the address and is turned away by the screen itself, correctly. The regulation library and the model status screen sit under Administration by their web address, and both admit an organisation administrator past their own check; the regulation library then refuses that same account on every save, because its real rule is narrower still.

Four screens, three audiences.

Screen	Who can open it	Who the underlying rule actually admits
The chat and the data question box	Staff and above	The same group. Exact match
The regulation library	An organisation administrator or higher, by the screen's own check	Your organisation's super admin only
The model status screen	Anyone signed in, by the screen's own check	Also anyone at all, signed in or not; nothing checks it on the server either

A day in the life

Tom opens the chat before his morning walkthrough and asks who is checked in right now. He gets a real number back, site by site, because that one figure comes straight from this morning's check-ins. He asks a follow-up about last month's attendance trend for a grant report, and the number that comes back is zero, worded as confidently as the first answer. Nothing on the screen tells him the two numbers were built two different ways.

Later, Denise Okafor pastes a regulation question about outdoor play into the same box. If your organisation has loaded the relevant DELACARE text, the answer is drawn from it; if nobody has, the assistant answers from its own general knowledge and is supposed to say so. Either way, it cannot tell her which paragraph of which document it used, because that detail is thrown away before it reaches the screen.

What PTRS does for you

• Strips out personal details before a question leaves the pipeline, for eight kinds of information: social security numbers, emails, phone numbers, street addresses, dates of birth, and children's, staff members' and guardians' names. Each is swapped for a placeholder and put back once the answer returns.
• Reads the answer again before you see it, watching for four of those eight kinds of information slipping through, including on an answer that is still arriving word by word.
• Decides what kind of question you asked, out of 23 kinds, and sends it to a stronger or a faster model at a temperature suited to the task.
• Pulls in your organisation's own live figures for many of those 23 kinds, on attendance, incidents, compliance, an organisation-wide summary or a pattern review, instead of answering from general knowledge alone.
• Answers regulation questions from documents your organisation has loaded, finding the five closest passages and handing them to the model under a source label.
• Limits how much any one person can ask for, twenty questions a minute and five answers arriving at once, so one person cannot run up an unreasonable cost on their own.

What stays your job

• Checking any number the assistant states in words, especially about attendance, against the screen that actually measures it. See below for which numbers to trust and which not to.
• Loading the regulation text yourself. Nobody is notified when a document is out of date; if your organisation never loads DELACARE or OCCL text, every regulation question is answered from the model's general training instead, with no label saying so.
• Repeating yourself. The assistant keeps nothing between messages. Ask a follow-up the way you would ask a stranger, with the subject named again.
• Deciding whether a real child's details may leave your own equipment at all, and setting the model status screen's language accordingly before anyone uses the chat for real questions.

What has been built

Screen	What it is	What it can do today
The chat	A conversation box with a message thread and a few suggested questions above it	Answers in words, and for about half of the 23 kinds of question, adds a chart, table or set of cards underneath
The data question box	A single plain-English box	Turns your question into an instruction it runs directly, showing you both the instruction and everything it found
The model status screen	Seven rows describing which model is in use and whether it can be reached	Read-only, refreshed automatically
The regulation library	A list of what has been loaded, with a form to add more	Loads a document into the same collection every regulation question draws from

What happens on its own

Nothing on this page runs on a fixed schedule; PTRS's timed jobs all belong to other modules. Four things still happen without being asked for:

• The first question after a restart pays a short delay while PTRS works out what each of the 23 kinds of question sounds like, then remembers that for as long as it keeps running.
• An answer that does not depend on the exact moment it was asked is remembered for five minutes, so an identical question asked twice in a row can come back instantly the second time.
• The model status screen checks whether the model can be reached once a minute while it is open. When the model is the outside service rather than your own equipment, that check is itself a small, real request to that service.
• Every question is limited automatically, so a burst of activity from one person is slowed down rather than left to run unchecked.

What it connects to

Module	The connection
Attendance	Almost every attendance figure the assistant can put into words comes from a summary table nothing currently fills in. The one exception is today's live headcount, which reads check-in records directly. See Attendance
Incidents	Three assistant actions duplicate work the Incidents module already does on its own screens, and none of the three is reachable from any screen today
Compliance	A compliance question is answered from each site's latest score and open alerts, averaged across sites with no weighting for size. See How the compliance score is calculated
Programs	Seven brainstorming actions for a program idea, an outline, goals, an assessment and a compliance check are counted and described here, though their two screens live under Programs
Administration	An administrator invites the people who can reach these screens, and only your organisation's super admin can add to the regulation library. See Administration
Dashboard	The regulation library's loading progress bar rides on the same live-update connection the home screen uses. See Dashboard

Not yet available

Following up on your own question

The conversation list and every message in it live only in your own browser. Reloading the page loses all of it, and the question PTRS sends carries no memory of anything you asked before. Asking "show me more" is answered as though it were the first thing you had ever typed.

A complete way of remembering a conversation exists inside PTRS and nothing in the product ever calls it. The model status screen nonetheless reports conversation memory as switched on, because that row only reads a setting, not whether anything actually uses it.

Trusting the data question screen on a copy of PTRS shared by more than one organisation

The data question screen checks that you belong to an organisation and then never uses that fact again. What it builds from your question runs directly, outside the boundary that keeps every other screen in PTRS inside your own organisation's records.

The safety check in front of it blocks anything that is not a plain read, blocks a short list of dangerous words, and is supposed to block anything outside an approved list of record types. The way the model is told to phrase its answer happens to slip past that last check entirely, so in practice nothing stops it from reading a record type outside the approved list either.

The rows that come back are shown exactly as they arrived: not re-checked for personal details, and not limited to your own organisation.

Naming which regulation document an answer came from

The retrieval step works out which document each passage came from, its section, and how close a match it found for all five passages, and none of that reaches the screen. What does reach the model is a numbered label for each passage, so an answer can say "per Source 3" with nothing on the screen able to tell you what Source 3 is.

Recording where an indexed regulation document came from

The loading form has an optional field for where the document came from, as a web address. It is collected and never saved anywhere.

Five of the chat's answer panels

Five of the chart, table and card layouts the chat is built to show are complete on the screen's side and never sent by the server, including a translation panel and an incident report layout.

Four suggested questions the chat offers

Translating a message, drafting a note to a guardian, sorting a meal description, and summarising attendance for a date each have their own, purpose-built action behind them, complete and working, and nothing on any screen calls them. The chat's own suggested questions name three of these by description; typing one into the chat routes to a general answer instead of the purpose-built one.

Where to go next

• Use the chat → Ask the assistant a question
• Load a regulation document → Index a regulation document
• See who can reach each screen → Who can do what with the AI Assistant
• See what each of the 23 kinds of question can and cannot answer → What the assistant can answer
• The question worth asking before anyone relies on this → How PII is protected in the AI pipeline
• Something is not working → AI troubleshooting