Appearance
Ask AI Documents
Upload your company's own documents — policies, procedures, permits, safety data sheets — and Ask AI can quote them in its answers, alongside what it already knows from your site's records.
Open Admin → Ask AI Documents. Everything on this page is private to your workspace: your documents are indexed into your workspace's own store, and no other workspace can retrieve them.
Uploading
Drop a file on the upload area, or click it to browse. Accepted types:
| Type | Notes |
|---|---|
| Scanned PDFs work. A scan has no digital text, so the text is read off the page image automatically — nothing to configure. | |
Word (.docx, .doc) | Paragraph text is indexed. |
| CSV | Each row becomes searchable content. |
| Markdown, JSON, plain text | Indexed as-is. |
Files can be up to 25 MB. Each workspace can hold up to 200 documents; if you reach the limit, remove a document you no longer need.
When the upload finishes you'll see how many sections were indexed. From that moment the document is quotable — ask a question about its contents and it appears among the answer's sources, cited by its filename.
Everyone in your workspace who can use Ask AI can get answers drawn from uploaded documents. Upload things your whole site is allowed to read.
Labels: organizing documents by source
Every upload can carry a category label — pick one next to the upload area, or type your own (OSHA, EPA, Company, SDS, Training, Images are offered as starters). Labels group documents by where they came from, and they are what make removing a whole category at once possible.
- The label is optional. A document without one shows under Unlabeled.
- Labels also power the ask page's Search sources row: once any label is in use, people can narrow a question to certain categories (see Ask AI).
- Change a document's label anytime with Label next to it; clearing it moves the document back to Unlabeled.
- Spelling case doesn't matter —
oshaandOSHAare the same label. - Replacing a document's file keeps its label unless you pick a new one while replacing.
The Category dropdown above the list filters it. With a label selected, Delete all in category removes every document under that label in one confirmed action — each one leaves the index exactly as a single Remove would, and a document whose removal fails stays listed so nothing quotable ever disappears from view. Past answers that already cited the documents are not affected (that is what Purging is for, one document at a time).
Unlabeled documents can only be removed one at a time, on purpose: "delete everything without a label" would delete whatever nobody got around to labeling.
Updating a document
Use Replace next to the document, then upload the new file. The new version takes the old one's place — Ask AI answers from the revision, and the old content is gone from the index. Uploading a revision without Replace indexes it as a second document, and answers may quote either copy, so Replace is the right way to publish a new revision.
Removing a document
Remove takes the document's content out of the index immediately — the next question cannot draw on it, no matter who asks. Two things to know:
- Answers people received before the removal are their own history and are not changed by it.
- If you upload the same file again later, its content is back in the index. Removal is not a ban on the document, just its removal today.
Seeing and editing what Ask AI can quote
View content next to a document shows its indexed pieces — everything Ask AI can draw on from that file, exactly as stored. It answers "what does the AI actually know from this document?" directly, instead of by asking test questions.
It is also how you remove part of a document. If one paragraph should never have been included — a salary figure in an otherwise-fine policy, a name in a report — paste the sensitive text into the find box: matching pieces filter into view. Tick them, and Remove selected takes exactly those pieces out of the index. The rest of the document stays answerable.
Three things to know:
- Nothing is ever removed by matching alone. The find box locates; you select. Scanned documents especially can have text that reads slightly differently than the original, so a match is a lead, not a verdict — read the piece before ticking it.
- Removing pieces changes what Ask AI can quote from now on. Answers people already received are untouched — if the content must also be swept from history and memory, that is what Purge below is for.
- Re-uploading the original file brings removed pieces back, because Replace and re-upload index the file as given. The durable fix is Replace with a properly redacted file; removing pieces here is the immediate one.
Purging a sensitive document
Purge is the escalation past Remove, for a document that should never have been uploaded — the wrong file, or one containing sensitive content. Where Remove stops the document being quotable from now on, Purge also sweeps up what already happened:
- Past answers that quoted the document are deleted from the Ask AI history of every person who received them. Only those answers — other conversations are untouched, and people's own questions are never deleted.
- Those same people's Ask AI memory is cleared entirely. Memory has no way to forget a single document, so the honest operation is a full reset for the people affected; their history and memory simply start fresh.
- A removal report is written to this page's removal log: what was purged, when it was uploaded, how many answers were scrubbed, how many memories were cleared — and whether any of that failed, stated rather than rounded away.
One limit is reported honestly instead of hidden. If your workspace was using a cloud model provider, every answer that drew on the document also sent that content to the provider, and nothing here can recall it — the report names the provider and the exposure window when that applies. With the local model ("data never leaves this server" in Ask AI Settings), nothing ever left, and a purge really is complete.
Purging cannot reach people's screens or their memories of what they read. It removes what the system holds; treat genuinely sensitive exposure as an incident too, the way you would any other disclosure.
What the upload does not do
The file itself is not stored — it is read, indexed, and discarded. What remains is the indexed text. That means Ask AI Documents is not a file cabinet: keep your originals where you keep them today, and treat this page as the list of what Ask AI has been given to read.
A watched folder is the other way round — the originals stay in the folder, which is what lets a deletion there remove the document from the index.
Watched folders
Uploading covers documents you add one at a time. A watched folder covers a directory that changes on its own — a shared drive of safety manuals, an export that a system writes nightly.
Your site gets one watched folder, and it covers the whole site: every person with Ask AI can find what is in it. There is no per-person folder and no per-department folder.
What it does that uploading cannot
A watched folder is re-read on a schedule and the index is made to match what is in the folder:
| in the folder | in Ask AI |
|---|---|
| a file is added | it is indexed on the next scan |
| a file is edited | it is re-read, replacing the old text |
| a file is deleted | it is dropped from the index |
That last row is the reason watched folders exist. Uploading can add and replace, but nothing tells Ask AI that a document on a shared drive was deleted — so without this, a folder slowly fills the index with documents that no longer exist anywhere else.
Turning it on
Open Admin Settings → Ask AI → Indexed Data. Under Watched folder:
- Choose how often to re-read it. Hourly suits a folder people edit through the day; daily suits a nightly export. You can change this later by stopping and starting the watch.
- Select Watch this folder.
That is the whole setup. There is no path to type — see Where the folder is below.
If the section instead says "Watched folders are not set up on this server", this deployment has no folder storage configured at all, and turning one on is a request to whoever runs your server.
Where the folder is
The screen shows the folder's location on the server — something like /data/tenants/8f3c…. That is not a path you can open from your desk, and it is not meant to be. Whoever runs your server maps a network share onto it, and that share is where your people put files.
Ask them which share writes to it. Until you know that, the folder is switched on and empty, which looks exactly like a broken folder.
If the folder does not exist on the server yet, Watch this folder says so instead of silently doing nothing — that message is what to forward.
Using it
Nothing more to configure. Put files in the share and they appear in answers after the next scan.
Indexed Data lists it as Watched folder with how many documents it currently holds and when it last synced. Re-sync now re-reads it immediately rather than waiting for the schedule — useful right after a bulk change.
Answers cite the file's own name, the same way an uploaded document does.
Stopping it
Stop watching ends the watch and removes what the folder put into Ask AI. The files themselves are untouched — this changes what Ask AI knows, not what is on the share.
Starting it again re-reads everything from scratch.
Why there is no path to type
The folder is read by the server, with the server's own permissions. If a path could be typed in, it could name the application's own files or a stored password, and the contents would come back through Ask AI looking like an ordinary answer.
So the location is not a choice. The server works out your site's folder from your site's own identity — which is also why your folder can never be another customer's, and theirs can never be yours.
Troubleshooting
"No readable text could be extracted from that file." The file contains no text the system can read — most often a scanned PDF whose pages are too poor to read, or an image-only file of something other than text. Re-scan at a higher quality, or export the document to PDF from its source application if you can.
"That file type is not supported." Only the types in the table above are accepted. For spreadsheets, export to CSV first.
A very long scanned PDF was only partly indexed. Reading scanned pages takes real work per page, so very long scans are read up to a page limit set by your server. Split the document, or ask your operator about the limit.