Add a support agent to your site

Brief an AI support agent (Gemma), drop one script tag on your site, and give your visitors a chat bubble that answers from your own knowledge — in your brand, around the clock.

Who this is for

You run a product or a site — a GATE-hosted project or anywhere else you can add a line of code — and you want visitors to get real answers without you staffing a chat window yourself. This isn't a bot that hands out a help-article link and hopes for the best: it reads from knowledge you give it, and it says so honestly when it doesn't know something.

Before you start

• Gemma turned on for your org. If your bot can't find it or the dashboard doesn't show a Gemma page, ask your org's admin to check tool access for Gemma.
• A site to put it on — a GATE-hosted project, or any other site you can add one script tag to (yours to edit, or hand off to whoever maintains it).
• A decision on what the agent is for. Each install picks one job at setup, and changing it later is a deliberate, rare move — not a toggle:
  • Frontend — answers anonymous visitors from your knowledge. Chat only, no logins, no tools.
  • Backend — helps signed-in users once your site tells it who they are, and can act through webhooks you control.
  • Report — takes structured bug reports and feedback and routes them to a queue, even if nobody's watching chat.
• Something for it to know. An FAQ, a help doc, a returns policy — anything in plain text or markdown. You can add more later, but starting with nothing means every answer is "I don't know."

Steps

1. Brief the agent. Decide how it should sound and where it should punt to a human — plain language, like briefing a new hire. Your bot can help you draft this: "Help me write a short, friendly persona for our support agent, and a line it should use when it doesn't know an answer." You'll paste the result into the install setup in the next step.

2. Create the install. This part happens in the dashboard, not chat — Gemma installs aren't something your bot can spin up yet. Open the Gemma page, start a new install, and step through name, mode, your persona text, a theme, and the site's address (it only answers requests from addresses you list). Finishing the wizard gives you a script tag tied to that install.

3. Put the script tag on your site. How, depends on where the site lives:

   • A GATE-hosted project: ask your bot — "Add the Gemma script tag to our site and deploy it." Your bot edits the page and pushes the change.
   • Any other site: hand the tag to whoever maintains that site's code. It's one line, and it can go anywhere on the page.

4. Point it at your knowledge. Ask your bot: "Upload our FAQ and shipping policy to the support agent's knowledge." Your bot files are stored as the agent's own reference material and read on every conversation, so answers stay consistent with what you actually wrote — not what the model assumes.

5. Test it before anyone else sees it. The dashboard has a "Try it now" preview — a sandboxed chat window that loads the real agent without touching your live site — so you can check tone and knowledge first. Once the script tag is live, ask your bot to spot-check it too: "Ask the support agent what our return window is."

What you'll see

• In the dashboard: the install's page — persona, theme, the script tag, allowed addresses, and (for Report mode) an incoming queue of structured reports.
• On your site: a chat bubble in the corner of the page, styled to your theme, with no visible trace of GATE in the markup.
• In chat: knowledge files you've handed over, and — for Report-mode installs — a summary when a new report comes in.

Troubleshooting

• The bubble never shows up. Confirm the script tag actually made it onto the live page (not just committed — deployed), and that the site's address is one you listed in setup. A mismatched address is silently refused, not an error you'll see in chat.
• The bubble loads but won't send a message. Usually the same cause — the page's address doesn't match what the install allows. Ask your bot to double-check the addresses on the install.
• It doesn't know things it should. Knowledge only updates when you hand it over — ask your bot what's currently loaded, and re-upload anything that's changed since setup.
• Backend mode doesn't recognize signed-in visitors. That needs your site to tell the agent who's logged in on every page load. If that wiring isn't in place, the agent falls back to treating everyone as anonymous — ask your bot (or whoever built the login flow) to confirm the hookup exists.
• Report mode: a report seems to have vanished. It hasn't — every report lands in the dashboard queue even if whatever system you route it to is down. Check there before assuming it was lost.

Going deeper

• Launch a landing page — get a site live first if you don't have one yet