Index a regulation document

Put the text of a DELACARE or OCCL document into the library, so a regulation question the assistant is asked draws its answer from that document instead of from the model's own general knowledge.

Who can do this

• Super admin

Two checks stand in front of this screen, and neither is the one that decides whether a save actually works. The wider one lets four roles reach the screens under Administration; the screen's own check narrows that to an organisation administrator or a super admin. An organisation administrator passes both, sees the loaded documents listed, sees the load form, and every load and delete they try fails, because the save behind all three checks a narrower rule the screen does not.

Before you start

Have the text ready. The form accepts a plain text file, or text pasted directly in. It does not accept a PDF: choose a file and PTRS will only offer plain text formats, and pasting the contents of a PDF in as text would load whatever unreadable characters that produces.

Pick a name you can live with. Everything else uses it: the row in the list, removing it, and loading it again later. It may only contain letters, numbers, hyphens and the mark _, and must be 200 characters or fewer. DELACARE-2024 is the example the form offers.

The library is shared by the whole installed copy of PTRS. A regulation document is deliberately not kept separate by organisation, and the search behind a regulation question adds no such separation either. What you load, every organisation on this installed copy of PTRS can draw answers from. What you remove, you remove for all of them.

Steps

1. Open the regulation library

The heading is Regulation Knowledge Base. The first card, Indexed Sources, lists what is already loaded with a count of passages and when it was last loaded; the second is the load form.

2. Fill in Source Name

Letters, numbers, hyphens and the mark _. If you choose a file first, the name is filled in from the file's own name, with anything else replaced by a hyphen.

3. Leave Source URL empty

The field exists and is sent along, and PTRS does not keep it anywhere. Filling it in records nothing.

4. Choose a text file, or paste the text

The text box shows a running character count. Index Document stays disabled below 100 characters, and PTRS enforces the same minimum on its own side.

5. Press Index Document

Checkpoint

A progress card appears above the form, naming the document and showing a count and a bar. The updates arrive over the same live connection the home screen uses, sent every ten passages processed.

The number the bar counts toward is an estimate, and it is always a little low, because the actual splitting keeps a short overlap between one passage and the next that the estimate does not account for. The bar can reach full before the work is actually finished. Worked out

6. Read the confirmation message

"Indexed N passages for '{name}'". This N is the real count, read back after the save completes, not the estimate the bar used.

Seen in PTRS

Checkpoint

A row for your source appears in Indexed Sources with that count and the current time.

Result

The document is split into passages of roughly 2,048 characters on sentence boundaries, with 256 characters carried over from one passage into the next, and each is turned into a numeric fingerprint and stored.

From then on, a question the assistant sorts as a regulation question retrieves the five closest passages across the whole library and hands their text to the model under a numbered source label.

The splitting step also looks for a DELACARE section reference inside each passage and keeps it. That reference does not reach the answer. See below.

Refreshing and removing a source

Loading the same source name again is safe, and is the intended way to update it. Every existing passage under that name is retired first, and the new set is written in its place.

Remove is the trash icon on a source's row. The confirmation states the passage count and says the material can be loaded again later; removing it retires the passages rather than erasing them outright. A removal applies across the whole installed copy of PTRS.

What this does not do

Tell you which document an answer came from

The retrieval step works out which document each retrieved passage came from, its section, and how close a match it found, for all five, and nothing else in PTRS reads any of that.

The model is given a numbered label for each passage, so it can write "per Source 3", and there is nowhere on any screen that says what Source 3 is.

Keep a document to one organisation

There is no field for it and no place it is kept. See Before you start.

Check that what you loaded is actually a regulation

The only checks are the 100-character minimum and the pattern the name must match. Any text at all can be loaded, and it will be retrieved and quoted as though it were a regulation.

Common errors

What you see	Why	What to do
A message saying you do not have permission to access this page	You hold a site director or regional director account. The wider check let you reach the address; the screen's own narrower check stopped you	Correct outcome, and stricter than it needs to be. The real rule wants a super admin
The screen loads, Indexed Sources is empty, and Index Document fails	You hold an organisation administrator account. You pass both checks on the screen itself and fail the rule behind all three saves	Ask your organisation's super admin
The progress bar sits at nothing and the confirmation eventually appears anyway	Progress is only sent every ten passages and once at the estimated total. A short document never reaches ten	Expected
The bar reaches full and keeps going	The number it counts toward is an estimate that is always a little below the real total. See step 5	Expected; the confirmation message has the real number
No progress appears at all, and the load still succeeds	The progress update is sent without waiting to confirm it arrived, so a lost connection to the live-update service is never noticed. Your browser may not currently be connected to it	Reload the page after the confirmation
A message saying the document text is too short to be a valid regulation document	Under 100 characters	Not applicable
A message saying the name must contain only letters, numbers, hyphens and the mark _	A space, a period or an accented character in the name	Not applicable
An unexpected failure that names the AI service, while the chat still works	Loading calls the model that turns text into a numeric fingerprint, which always runs on your own equipment regardless of which model answers chat questions. If that is not running, loading fails even while chat succeeds	Check the fingerprinting model's status dot on the model status screen

Related

• AI Assistant overview
• What the assistant can answer
• Who can do what with the AI Assistant
• AI troubleshooting