This guide explains each part of NEXUS in plain language. Open it any time with the ? button beside a page's title — it brings you straight to that page's section.

Getting started

NEXUS reads guest emails from the hotel's inbox, writes a suggested reply using the hotel's own documents, and either sends it automatically (when it is very confident) or places it in the Review queue for a person to check. Nothing about the AI is a black box here: every draft shows the exact document excerpts it relied on.

Sign in with your work email and password. If your organisation requires two-factor authentication, you will be guided to set it up on first sign-in — you need an authenticator app such as Google Authenticator or Microsoft Authenticator on your phone. After five wrong password attempts, sign-in locks for a short period; simply wait and try again.

Mailbox

The Mailbox lists every guest conversation, newest first. Click a conversation to read the full exchange — guest messages, the AI's replies, and any attachments with the text the system extracted from them.

What you may see on a conversation

• Pending — a draft reply is waiting in the Review queue.
• Auto-sent — the AI was confident enough to send without review (this only happens if auto-send is switched on in Settings).
• Needs human — the AI decided a person must handle this: complaints seeking compensation, billing disputes, legal or medical matters, or questions the documents cannot answer.

Ticking several conversations lets you delete them together; the 🗑 on a row deletes just that conversation. Deleting removes its emails, drafts and follow-ups.

Review queue

Every draft the AI is not allowed to send by itself waits here. For each one you see the guest's message, the suggested reply, a confidence score, and the grounding — the exact knowledge base excerpts the answer came from, so you can verify rather than trust.

Your three choices

• Approve — sends the reply as-is (or with your edits; the preview always shows your edited text).
• Edit then approve — fix wording or facts first. Your edits are measured, which helps show how "send-ready" the AI's drafts are over time.
• Reject — nothing is sent; handle the guest by other means.

Understanding confidence

Confidence is the AI's own honest estimate of how well the evidence supports its answer. Settings defines two lines: above the auto-send threshold a draft may go out by itself (if auto-send is on); below the review threshold a draft is marked as needing extra care. An optional independent "judge" pass can double-check drafts and lower confidence when the answer is not properly supported.

Follow-ups

When a reply promises an internal action — "our team will confirm your late checkout" — that promise becomes a follow-up task here, so nothing said to a guest is silently forgotten. Each item shows the guest, the promise, and its source conversation. Mark it done once the action is taken; approving a resolution in the Review queue closes the matching follow-up automatically. Teams can be configured to receive these by notification email or webhook.

Knowledge base

The knowledge base is the set of documents the AI is allowed to answer from — and the single biggest factor in answer quality. Upload PDFs, Word files, spreadsheets or text; scanned or image pages are read by AI vision once and remembered.

Uploading and versions

• Uploading a file with the same name replaces the old one — the previous copy is archived automatically (nothing is ever lost) and the document's version number increases.
• Uploading something that looks like a new version ("Facilities v1.1") triggers a choice: replace the original (archived), keep both (with a warning — two live versions can contradict each other), or skip.
• Deleting a document also archives it first.

Visibility and property tags

Public documents can be quoted to guests. Internal documents inform the AI's decisions but are never quoted. Inside public documents, any line beginning with "Internal info" is removed before the AI sees it. A property tag ties a document to one hotel, so questions about Hotel A are answered from Hotel A's documents.

Test a query

Type a question a guest might ask and see exactly which document excerpts the system retrieves — before any guest does. The property dropdown scopes the test to one hotel, and Compare modes shows the same question under each retrieval method. Use this after every significant upload.

Reindex and Force re-extract

Reindex rebuilds the search index from the documents — quick, safe, run it after uploads. Force re-extract additionally re-reads every file from scratch, including AI vision on image pages — slower and it re-spends vision processing, so use it only when documents were extracted incorrectly.

AI agents

An agent is a personality-and-scope package: its name, tone and language style, which knowledge folder it answers from, and which hotels it covers. One agent per brand or business line is typical. The agent used for a conversation stays with that conversation, so a guest never experiences a mid-thread personality change. Every agent obeys a set of built-in, non-editable safety rules — internal information, sarcasm handling, conflicting documents — regardless of how it is configured here.

Teams

Teams group the humans: Front Office, Reservations, Engineering. Incoming emails are routed to a team (by the AI's classification, or pinned by mailbox — see Mailboxes), and that team's members review the drafts. Each team can have a notification email or webhook for escalations and follow-ups.

Users & roles

Every person gets a user account with a role, and a role is simply a list of permissions — view mail, approve drafts, manage the knowledge base, manage users, and so on. Two ideas matter:

• What you cannot do, you do not see. Missing buttons or panels are the permission system working, not the page breaking.
• Team scoping. A user limited to certain teams sees only those teams' conversations, drafts and follow-ups.

Sensitive abilities are deliberately separate permissions: deleting conversations, clearing cost history, resetting platform data, and even viewing which AI model and software build is running.

Mailboxes (Settings → Mailboxes)

NEXUS can monitor several inboxes at once — reservations@, spa@, events@ — each with its own preferred AI agent, and optionally pinned to one team. Two rules to remember:

• Replies always send from the inbox that received the email. A guest who wrote to spa@ gets an answer from spa@.
• Passwords are write-only. The app password you enter is stored encrypted and never displayed again; when editing, leave the password field blank to keep the existing one.

If one mailbox has a wrong password, it logs a warning in Activity and the other mailboxes carry on unaffected. Emails between your own managed mailboxes are never auto-answered, which prevents reply loops.

Email templates

Templates control the visual dress of outgoing email — header, colours, footer, disclaimers — while the AI writes the content inside. Use the test-send to check rendering in a real inbox before adopting a template.

Reports

Reports answer two questions: is the AI helping, and what does it cost.

• Volume and outcomes — emails handled, drafts auto-sent vs reviewed vs rejected.
• Avg draft modified — how much humans edit approved drafts; lower means drafts arrive closer to send-ready.
• First response (auto / human) — how quickly guests get an answer on each path.
• Costs — every AI call is recorded with its purpose (drafting, retrieval ranking, quality judging, vision, query rewriting), so cost per feature is visible and auditable. One-time knowledge base indexing cost is shown separately from day-to-day running cost.

Date filters at the top narrow every figure. The date boxes inside the reset card belong to the reset function and appear only for accounts allowed to reset data.

Settings

Settings is for how the system behaves (personal matters live in My account). The consequential ones:

• Auto-send and its threshold — whether high-confidence drafts go out without review, and how high "high" is. Start with auto-send off; switch it on once the Review queue has earned your trust.
• Review threshold — below this confidence a draft is flagged for extra care.
• Retrieval settings — how the knowledge base is searched. The default (hybrid with reranking) is the recommended, field-tested mode.
• AI query rewriting — optional: an extra small AI pass that rewrites multi-question emails for better search. Off by default; enabling it shows the cost implication first.
• Data redaction — personal data (IDs, card numbers, phone numbers) is masked before any AI reads it.
• 2FA policy — make two-factor optional or compulsory for all users.
• Mailboxes — see above.

My account

Your personal page: display name, profile photo, password, and two-factor authentication.

• Photo — PNG/JPEG/WebP up to 3 MB; it is resized and re-encoded for safety.
• Password — changing it signs out every other device where you were logged in (your current session continues). Minimum 8 characters.
• Two-factor — scan the QR with your authenticator app, confirm one 6-digit code, done. Turning it off also requires a valid code. If you lose your phone, an administrator must clear 2FA on your account so you can re-enrol.

Activity

The system's diary: sign-ins and lockouts, document uploads and replacements, mailbox problems and recoveries, configuration errors, migrations. When something seems off, look here first — the design principle throughout NEXUS is that failures announce themselves rather than hide.

FAQ

Why didn't the AI answer a question the documents do cover?

Usually retrieval: test the exact question in Test a query. If the right excerpt appears there but not in the draft's grounding, note the email and report it — that comparison is exactly what the engineering team needs.

Why does a document show "v2 · updated …"?

It was replaced at some point; the old copy is archived, not gone. The AI is told each excerpt's date and prefers newer documents when two disagree — and flags the conflict for review.

Why did a guest asking for an internal code get refused?

By design. Content marked internal is stripped or refused with a polite public alternative. If a genuine public fact was refused, check whether the source document marks it "Internal info".

Why can't I see a setting or button my colleague has?

Permissions. Ask an administrator to check your role on Users & roles.

The AI replied "a colleague will follow up" — is that a failure?

No — that is the escalation path working. The alternative would be a made-up answer, which NEXUS is built to never give.