AI Review & KB Review

What you are doing (two jobs)

Nexus keeps two operator workflows separate on purpose. Mixing them up is the most common confusion.

	AI Review	KB Review
Goal	Handle live tickets with an AI draft reply	Curate durable product knowledge from good staff answers
When	After AI screening runs on a ticket (new or re-screened)	On closed/historical tickets, imports, or any ticket you mark for knowledge work
Board view	AI review	KB review
Ticket tab	AI draft	KB review
Outcome	Accept, edit, send, or reject a draft public reply	Grade the reply, assign product, promote a gold answer into knowledge
Independence	KB review does not depend on AI draft review. You can do either alone.

The closed loop

1. Ticket arrives (email, public intake, SPA create, or historical import).
2. AI screening classifies the ticket (category, confidence, product hints) and may draft a reply using active product knowledge.
3. AI Review — a human accepts, edits, sends, or rejects the draft. Optional quality grade and notes train judgment over time.
4. KB Review — on good historical or resolved tickets, staff mark the best public reply, score quality, and promote it as knowledge (usually inactive until curated).
5. Activate knowledge on Tickets → AI & knowledge so the next screening pass can cite better material.

AI Review — step by step

Use this when the board shows tickets that need a human decision on an AI draft.

1. Open the SPA and go to Tickets.
2. Switch the board view to AI review. These are tickets where AI has produced a screening result that still needs a human outcome (badge: AI needs review).
3. Open a ticket. The default tab is often AI draft when review is pending.
4. Read the screening summary:
   • Category and confidence (e.g. noise, help, or other tenant categories)
   • Policy decision and recommended action
   • Product resolution (which product knowledge was used)
   • Citations from knowledge sources the model relied on
   • Flags and optional reasoning (expand if present)
5. Read the Draft reply. Click Edit draft if you need to fix tone, facts, or links before sending.
6. Optionally set Quality grade (1–5) and short review notes (why accept / edit / reject).
7. Choose the ticket status with reply (dropdown) if you are sending—status is not forced to closed.
8. Pick an action (see below). Use Re-screen if inbound context changed or you want a fresh draft after knowledge updates.

AI Review actions

Action	What it does
Send draft / Send edited	Posts the (possibly edited) draft as a public ticket reply and records the human outcome. Applies the status you selected with the reply.
Send and close	Same as send, then leaves ticket detail (returns to the board). Status still comes from the dropdown—not an automatic hard-close unless you chose a closed status.
Accept (no send)	Marks the draft reviewed/accepted without sending a public reply (useful when you handle the customer another way or will reply manually).
Reject draft	Discards the draft as unusable. Ticket stays open for a human reply; you can re-screen later.
Re-screen / Queue AI screen	Enqueues a new screening job (after more messages, knowledge changes, or first-time screen).

KB Review — step by step

Use this to mine good support answers into product knowledge. Common sources: closed tickets, historical imports (e.g. osTicket), or any ticket tagged for knowledge work.

1. Open Tickets and switch the board view to KB review. Queue is tickets with KB status Pending or In progress.
2. Open a ticket → tab KB review (labeled “Response / KB review”).
3. If the ticket was imported, note the external system / id banner for context.
4. Set fields:
   • KB status — Pending → In progress while working; Done when finished; Skipped if not useful for knowledge.
   • Product — required before promoting knowledge (which product this answer belongs to).
   • Response quality (1–5) — how good was the staff public reply for reuse?
   • Notes — short curation notes for other reviewers.
   • Gold public reply — pick the best public message on the thread.
5. Click Save (or save with a status change) so the board queue updates.
6. When you have a gold reply and product, promote that message to knowledge (next section).
7. Mark the ticket Done (or bulk mark done on the board when you only needed triage).

Promote gold answers into knowledge

1. On the KB review tab, ensure a product is assigned.
2. Select the gold public message (or use promote on that message).
3. Promote creates a product knowledge source—usually inactive with tags such as pending-curation and from-ticket-review so nothing goes live by accident.
4. Go to Tickets → AI & knowledge.
5. Find inactive snippets pending curation. Edit title/body/tags if needed, then set Active.
6. Active knowledge is what AI screening may cite on the next pass (when the KB loop is enabled).

You can also add knowledge manually on the same page as Url, Snippet, or WebsiteRoot sources—not only from ticket promotion.

Configure AI & knowledge

Path in the SPA: Tickets → AI & knowledge (tenant AI screening policy + knowledge library).

Setting	Meaning
Screening enabled	Master switch for AI screening on this tenant.
Auto-send when policy allows	Allow automatic public send for safe high-confidence cases.
Auto-send min confidence	Threshold (0–1); higher is more conservative.
Auto-send categories	Category keys allowed to auto-send (e.g. noise,help).
Require citations for help	Help-class auto-send needs knowledge citations.
Rescreen on inbound	Queue a new screen when the customer writes again.
KB loop	Allow screening to use (and benefit from) curated product knowledge.

Platform hosts also run the in-process worker (Tickets:AiScreeningWorker) in Heuristic or Llm mode. LLM mode needs an API key (e.g. XAI_API_KEY or platform settings). Global admins can override worker settings under Admin → Platform settings.

Bulk board actions

On the tickets board, select multiple tickets (checkbox column), then use the bulk bar:

• AI review view: bulk Accept, Reject, or Send drafts (send confirms first).
• KB review view: bulk Mark done or Skip; optionally set product with the bulk product control.
• Standard bulk status / priority / assignee / product apply on any view.

Bulk AI send uses the stored draft as a public reply—spot-check a few tickets before large batches.

Operator tips

• Start the day on AI review for live customer wait time; use KB review for deeper curation blocks.
• Never promote secrets—strip passwords, personal data, and one-off account details before activating knowledge.
• Score honestly—quality grades and reject reasons help tune policy; rubber-stamping 5s hides weak drafts.
• Product first—KB promote requires a product so knowledge stays tenant- and product-scoped.
• Re-screen after KB activation if you want an immediate better draft on a stuck ticket.
• Imports: historical closed tickets often land as KB Pending for staff curation (e.g. after an osTicket historical import).