Docs · Quickstart
Live in 5 minutes
By the end of this guide you'll have a widget answering from your knowledge base and an agent that has completed its first task. Code is needed only for the embed step (copy-paste).
What is SemAgent?
- RAG Widget — a chat widget that answers visitors from a knowledge base built from your site and documents, citing sources.
- AI Agent — runs tasks across email, calendar, Drive, databases and Shopify, with your approval.
They share one knowledge base: whatever you teach the widget, the agent knows too.
Quickstart: install the widget (5 min)
- Create a free account at semagent.ai/register (no credit card).
- In the dashboard go to Knowledge Base → Add Source, enter your site URL and start the crawl. Your sitemap is found automatically; forums are auto-detected.
- Get your tenant ID from Widget → Install, then paste this before your site's closing
</body>:
<script src="https://semagent.ai/static/widget.js" data-tenant-id="your-tenant-slug" defer></script>
Reload your site — the widget bubble appears bottom-right. Ask something your site answers → you get a sourced reply. Ask something it doesn't → the widget says so instead of making things up. That's strict-KB grounding working.
Quickstart: your first agent task
- Open the Agent tab. Under Settings → Autonomy, stay on Supervised while you learn: reads and drafts run automatically, sends and writes queue for your approval.
- Type a first task:
Summarize our return and shipping policy from the knowledge base into a short customer-ready document.
The agent uses kb_search and doc_generate — both read/draft class, so no approval needed. Then try "email it to me": sending is an action, so an approval card appears. Approve it, and you've seen the heart of the agent security model.
Core concepts
- Tenant: your isolated workspace — knowledge base, widget, agent and team live here.
- Strict-KB grounding: the widget answers only from your knowledge base, citing sources; below the threshold it says "I don't know".
- T1 / T2 tools: T1 = read/draft (side-effect-free), T2 = write/send (approval-gated).
- Autonomy level: how much risk the agent may take without asking (Supervised → Autonomous, 4 levels).
- Connector: a secured link to Google, Microsoft, Notion, your database or your MCP server; credentials live in an encrypted vault.