🤖AI Assistant

The AI Assistant summarises tickets for staff, takes a first pass at a member's problem before staff arrive, routes new tickets to the right category, lets a keyword answer in its own words, answers members' questions on demand with /ai ask, and gives your staff four tools: a guide to FiveUnity itself, reply drafts, channel catch-ups and a writer for announcements. It runs on your own free Google Gemini key.

#Built for FiveUnity

Out of the box, every AI feature knows what it is part of. Member-facing answers (the first responder, /ai ask and AI key phrases) introduce themselves as FiveUnity's assistant, understand a FiveM roleplay community's vocabulary, never make a staff decision such as a ban, refund or whitelist approval, and send anything that needs a human to a ticket. The first responder is told which kind of ticket the member opened and asks for the details staff usually need. The staff tools know the /ticket workflow, ticket routing knows how FiveM support usually splits, and summaries keep in-game names, ids, amounts and evidence exactly as written. The assistant also reads your FiveUnity setup (see Use this server's FiveUnity setup below), so it can point members at the right ticket panel.

None of this replaces your own settings: Server information, Personalization and the guidance for each tool still come first.

#Set it up

1. Pick a free provider (or a recommended mode):
   • Accurate (recommended for summaries, routing and drafts): Google Gemini. Free AI Studio key, no billing account.
   • Fast (recommended for first replies and /ai ask under load): Groq. OpenAI-compatible, very low latency.
   • Optional: OpenRouter for a free-model router (openrouter/free) when you want one key that can reach several free models.
2. Paste the key into the provider’s API key field, pick a Model, and press Test connection. A successful test reports the model and how long it took.
3. Save, then fill in What the AI knows; without it the model has nothing to answer from.
4. Under Data-use consent, press Review and accept. Ticket summaries, the first responder, escalation, reply drafts and catch-ups stay locked until you do.
5. Switch on the features you want.

Your key is stored encrypted and never shown again. Type a new one to replace it, or press Remove keys: every AI feature stops working immediately.

Every step above can also be done from Discord with /ai panel, described below.

#Providers

Dashboard Fast / Accurate buttons set provider and model together. Switching provider clears model detection and may need a new key for that vendor.

#The /ai panel control panel

/ai panel (or /setup → AI Assistant) opens everything about the assistant in one private message: connection, consent, features, knowledge and usage, with these controls.

/ai status is the read-only version for staff: whether the assistant is ready, the model, consent, today's usage and which features are on. /ai panel needs the setup permission; the rest of /ai needs ai.

#Google Gemini

Only Flash models are offered; they are the ones on Google's free tier. If a request fails, the last error is shown at the top of the panel, so you can tell a bad key from a quota block.

FiveUnity remembers which key in the pool last worked, including across bot restarts. A backup is tried only when Google reports a rate/quota limit; other errors are returned normally so a bad request cannot silently consume every key.

#Data-use consent

Any feature that sends member-written content to Google (ticket summaries, the first responder, automatic escalation, reply drafts, channel catch-ups, AI key phrases and AI log answers) is locked until an admin accepts the data-use terms, both in the dashboard and on the bot side. Accept on the dashboard, or in Discord with /ai panel → Consent, which shows the same notice. It says:

Content you send to Gemini on Google's free tier is used by Google to improve their products. Ticket transcripts routinely contain member names, ban appeals, and payment references. Only accept if you are comfortable with that, and tell your members in your own rules.

Accepting records who accepted and when, and the panel shows both afterwards. Nothing else about the decision is stored. See Privacy and your data.

#What the AI knows

The assistant answers only from what you give it. There are several sources and you can combine them.

Refresh knowledge rebuilds the block now and reports how many characters came from which sources. Otherwise it is rebuilt every 15 minutes, and whenever you save this page.

#Ticket summaries

Condenses a ticket into Summary, Outcome and Action items that staff can read in seconds.

Staff can also summarise on demand with /ticket summary inside an open ticket; the answer comes back privately. They can add question: to get a direct answer alongside the normal summary. It is granted through its own permission node, so you can give it to seniors without the rest of the ticket commands. See Permissions.

Summaries are for staff only. They go to a staff channel and are never DMed to the member.

#Ticket first responder

Answers the member's first real message in a new ticket, once per ticket, before staff arrive.

It only replies to the member who opened the ticket, never to staff or anyone added later, and it ignores very short messages so a bare "hi" does not spend your quota. The reply is an embed footed AI first response · staff will follow up. If it cannot be produced (quota, an outage, a missing key) the ticket carries on silently and it can try again on the next message.

The assistant is instructed never to claim to be a human or a staff member, never to invent rules or prices, and to say plainly when your server information does not cover the question. It cannot ping @everyone, @here or any role.

#Automatic ticket escalation

Reads the opener's first detailed message in a brand-new ticket, compares it with your knowledge base and category descriptions, then moves the ticket if the match is confident enough. It only ever acts on new tickets: once the routing window has passed or a staff member has posted, the ticket stays where it is.

When it routes

Where it moves tickets

Routing reads the opener's messages in the ticket (not only the latest one), uses the same daily quota as other AI features, and records the move like a normal escalation. With Auto-rename on, the model's short label is preferred over the type naming scheme. Rename is part of the same Gemini request as routing, not a second one. When the first responder is also on, both features share one request on that opener message (route, optional rename, and the first reply together), which keeps a free-tier key (~10 requests per minute per Google Cloud project) from being spent twice. A confident move, a clear "keep it here", or a below-threshold confidence counts as the one decision for that ticket; a stay-put decision can still auto-rename. If the model is unavailable, over budget, or returns unusable output, the ticket stays where it is and the next detailed opener message can try again, but only while the ticket is still inside the routing window and no staff member has posted. The bot also checks again just before moving: if staff replied, claimed or moved the ticket while the model was working, it leaves the ticket alone.

Staff can always route manually with /ticket escalate, or force a classify-and-move pass with /ai escalate. /ai escalate ignores the routing window, the staff-reply and claim switches, and Route tickets opened as, and it always posts the move and follows the Tickets page ping setting, like /ticket escalate. Deterministic automations are available separately under each ticket panel's Automations editor.

#AI key phrases

Lets a keyword trigger a generated answer instead of a fixed one.

Switch AI key phrases enabled on here, then set an individual keyword's mode to AI on Modules and utilities. An exact keyword match with a written answer always beats an AI keyword, so your fixed replies keep working and cost nothing. If AI is off, over budget or failing, an AI keyword falls back to its written response rather than going silent.

The member's message is what gets sent to Google, so this needs the data-use consent as well: until it is accepted, an AI keyword answers with its written response and both panels say it is waiting.

#Ask the assistant (/ai ask)

Members can ask the assistant a question on demand and get a private answer from the server information: "how do I get whitelisted", "what are the rules on RDM". It follows the same rules as the first responder, never claims to be staff, and points at a human for anything the knowledge does not cover.

The per-member cooldown and daily share under Usage limits apply to it, and you can restrict or disable the command per channel on the Slash commands page as well.

#Staff AI tools

Five /ai subcommands for staff. Each is one request, each answers privately (except /ai escalate, which posts the move in the ticket), and all of them are granted through the ai permission node (or a single subcommand such as ai.draft). They are on by default once the assistant is on, except escalation which follows the Automatic ticket escalation switch.

Full option tables are on AI commands.

#Usage limits

These keep a server inside Google's free tier. Lower is safer.

There is also a short per-channel cooldown, and if Google reports that you are out of quota the assistant rests and resumes automatically; the dashboard and /ai panel both show when it is resting and when it resumes, and a backup key is tried straight away.

#Personalization

Presets that apply to every AI reply on this server. In Discord they live under /ai panel → Personalize.

Guidance for one tool at a time lives with that tool: First-responder guidance and Response examples under What the AI knows, Drafting guidance under Staff AI tools, Routing guidance and Naming examples under Automatic ticket escalation, and the log-answer guidance and default answer shape on the Log Lookup page.

#AI-assisted setup

The last section of the page lists every module that can be configured from a plain-English description, with a link to its Set up with AI entry point and the state of the last plan you made there. Each flow uses this page's Gemini key and counts against the same daily budget: the assistant drafts a plan, FiveUnity checks every name against your server and asks when one could mean several things, you review the preview and the full list of changes, and nothing is built until you approve it.

Both flows offer section backups on the page: save ticket config or the channel layout before you apply, then restore if the result is wrong. Channel setup is create-only: it will not edit, move, delete or reorder channels that already exist, create roles, or change role permissions. Applying a channel plan also requires typing CREATE CHANNELS.

#Fixing an existing ticket setup

The ticket assistant is built for repair work as much as for new builds. It reads your current panels (including whether each panel's Discord message is still alive and has working buttons), so you can paste a symptom ("the dropdown does nothing", "staff cannot see app tickets") and it answers with a diagnosis and the smallest plan that fixes it, not a rebuild.

How edits behave, and how they stay safe:

• Edits keep what they do not touch. A plan that updates a panel leaves its unmentioned types, embed text, roles and channel alone. New types are added next to the existing ones; a type is only ever removed when you asked for a replacement list, and then the review screen warns about every type that would go.
• Nothing disappears silently. The review screen shows a field-level list of what changes on each existing panel, and a plan that deletes panels or categories you never asked to delete is refused outright.
• Name drift cannot duplicate a panel. If the plan's name is a near miss of an existing one, you get a warning instead of a second panel, and "rename X to Y" updates X in place rather than creating Y next to it.
• Your choices are remembered. Hand edits on the plan, answers to the assistant's questions, and any warnings from the previous version are all fed back to the model on the next request, so it stops repeating mistakes.
• Text answers work. When the assistant asks an advice question, you answer in words in the "Your pick" panel and it folds the answer into the next plan version.

#Related

• AI commands: every /ai subcommand and /ai ask, with their options.
• Staff logs: the staff log finder, which uses the same assistant for its written answers.
• Tickets: the ticket lifecycle the summaries and first responder plug into.
• Plans and billing: upgrading to Pro.
• Privacy and your data: what leaves your server and what does not.