Screen: /en/account/answers (shared with Support AI) · Counts: documents stored — a standing total, not a monthly allowance
Knowledge is the same library as Support AI, pointed inward: your own team asking questions of documents you control, with the answer quoted exactly and its source named.
How it differs from Support AI
They share one screen and one library. The difference is who asks and how it is counted.
| Support AI | Knowledge | |
|---|---|---|
| Who asks | Your customers, through the public question box | Your team |
| Counted | Questions asked, per month | Documents stored, as a standing total |
| Allowance resets | On the 1st of each month | Never |
The standing total is the thing to understand. A library is something you keep, not something you consume. Your plan allows a number of stored documents; it does not reset monthly, and the billing screen says so rather than showing a reset date. Withdrawing a document frees its room immediately.
Managing documents
Each source shows its status, how many answers it holds, its link and its version.
- Withdraw — stops it answering at once. Use this the moment a policy changes: a wrong answer
given confidently is worse than no answer.
- Mark stale — it keeps answering, but is flagged as needing review.
- Replace what this document says — stores a new version and keeps a fingerprint of the old one,
so there is a record that it changed and when.
Who can see which document
This is the control Knowledge exists for. Each document carries an audience:
| Audience | Who can be answered from it |
|---|---|
| Anyone | Everyone, including your public question box |
| Members of this workspace | Your own staff only |
| Owners and admins only | Not general members |
It is enforced where passages are searched, not where lists are drawn, so a restricted document is unreachable rather than merely hidden — the passage cannot come back to someone who should not have it. Narrowing takes effect immediately.
Existing documents are Anyone, because that is what they were before audiences existed and silently narrowing them would have changed what a live question box replies to.
Why there are no embeddings
Retrieval is explainable word overlap, deliberately. You can always see why a passage was returned and what its match score was. An embedding model would answer more questions and make "why did it say that?" unanswerable — which, for a document your staff act on, is the wrong trade.
Two different measures are used: a question-and-answer pair is matched symmetrically, while an uploaded document's passage is scored on how much of the question it covers. This is why "Do you ship to Egypt" does not come back with the Jordan delivery policy.
What it does not do yet
Reading a scanned PDF, which is a picture of a page with no text in it, and syncing a Drive or shared folder.