AI Knowledge Assistant
AI That Answers From Your Playbook — Or Not At All
“怎么处理家长投诉?” Your team asks in their own language; the assistant answers from your approved SOPs, policies and courses — with citations. Trust is the feature.
How Closed Retrieval Works
The assistant is powerful because of what it refuses to do: it never reaches outside your approved, permission-scoped content.
Upload And Approve
Your SOPs, policies and courses go through normal review. Nothing enters the AI's knowledge until a person has approved it.
Indexed With Permissions
Approved content is indexed together with its visibility scope — who may see it is part of the index, not an afterthought.
Ask, Get A Cited Answer
Staff ask in plain language, in English or 中文. The answer quotes only what that person may see, with a source card naming the document and version.
The Journey Of One Question
Five hops, a few seconds — and every one of them is a control point, not just a step.
Someone Asks
In English or 中文, in their own words — “怎么处理家长投诉?” works as well as a keyword.
Retrieval, Workspace-Only
The search runs inside your approved content. The open internet and other workspaces don't exist to it.
Permission Filter
Passages the asker isn't allowed to see are dropped before the AI ever reads them.
Answer Composed
The reply is grounded in the surviving passages — your words, your procedures, nothing invented.
Sources Attached
Every answer carries source cards: document, version, last updated. One tap opens the original.
And when nothing survives step 3? The assistant says there's no approved source and points to your admin — the unanswered question lands in the admin's gap list, so missing SOPs become a to-write list instead of a guess.
Five Guarantees, Engineered In
Every Answer Cites Its Source
Answers link to the exact SOP, policy or lesson — with version and last-updated date. Your team can always check the original.
Permissions Apply To AI Too
If a member can't see a course, the AI can't quote it to them. Retrieval runs through the same visibility engine as the rest of the platform — verified by automated tests.
No Source, No Answer
When approved knowledge doesn't cover a question, the assistant says so and points to your admin. It never invents company policy.
Injection-Hardened
Uploaded documents are treated as data, never as instructions. Suspicious instruction-like content is excluded at ingestion and audited.
Continuously Evaluated
A golden-question test suite checks citation accuracy and permission isolation after every change to models or content pipelines.
Practice Mode
Roleplay The Hard Conversation Before It's Real
Reading the SOP is level one. Saying the words to someone pushing back is a different skill — so the assistant plays the other side. Pick a scenario, set the difficulty, and rehearse until the score says ready.
- Scenarios built from your own SOPs — refunds, complaints, objections
- The AI plays the counterpart and pushes back like a real person
- Sessions scored on your rubric, with concrete feedback per criterion
“The studio down the road quoted 30% less for the same package. Give me one reason I shouldn't walk.”
Beyond Questions And Answers
The same trusted foundation powers drafting and upkeep — always as drafts a person approves.
AI-Drafted Courses & SOPs
Point the AI at your files and get a course outline, lesson summaries and a quiz draft in minutes. Every draft waits for human approval before anyone learns from it.
Auto-Tagging On Upload
New content is tagged automatically so the library stays searchable as it grows — admins can correct tags anytime, and corrections win.
Workshop Summaries
After a workshop, the AI turns attendance notes and reflections into a draft knowledge article the facilitator can approve and publish.
AI Questions, Answered Directly
What content can the assistant draw on?
Only content that exists in your workspace and has been approved: published lessons, current SOP versions and policies. Drafts, archived content and other companies' workspaces are never in scope.
What happens when there is no approved answer?
The assistant says it cannot find an approved source and points the person to their admin — it never fills the gap with a guess. Unanswered questions are visible to admins, so missing SOPs become a to-write list.
Can one company's questions or answers reach another company?
No. Retrieval runs inside your workspace only, and isolation is enforced at the data layer everywhere content surfaces — the security page shows the architecture.
Is our content used to train AI models?
No. Your content is retrieved at question time to ground the answer and is not used to train models. Removing content from your workspace removes it from retrieval.