Every feature, fully spelled out.
The short version: a curated scenarios layer for the questions you know about, an on-device AI layer for everything else, and a one-line embed so it ships wherever a script tag goes. Below: the long version.
Two layers of intelligence
Every visitor question is first compared against your curated scenarios using a tiny embedding model (cached in the browser forever). If the similarity score clears 82 %, that scenario fires in around 50 ms — exact wording, every time. If not, the on-device LLM takes over with RAG over your uploaded documents.
- Scenarios layer: instant, exact, no hallucinations, works on iOS.
- AI fallback: on-device Qwen3 0.6B (WebGPU, WASM fallback) with retrieval over your sources. English only at launch.
- Configurable threshold: tune the match cutoff per assistant.
Curated triggers · ~50 ms · iOS-friendly
Qwen3 0.6B · RAG · runs in the browser
Sources — files or your existing site
Drop in PDFs, Word documents, spreadsheets, Markdown, CSV, plain text. Or point Zupport.chat at your existing website: it crawls the pages, paginates the results, and lets you hand-pick exactly which subpages should become knowledge.
- File types: PDF, DOCX, MD, CSV, XLSX, TXT.
- Website crawl: point-and-click subpage selection with pagination.
- Manual subpages: add specific URLs that crawls miss.
- Realtime updates: the dashboard streams changes via Firestore.
Scenarios with AI on both sides
Write the questions your customers ask, or click Suggest question and let the AI propose them from your sources. Either way, the AI also drafts the answer for you — you skim, tweak, save.
- Suggest question: AI proposes scenarios from your knowledge base.
- Draft answer: AI writes the response, grounded in your documents.
- Multiple triggers per scenario: capture wording variations.
- Reorder & group so the most-asked questions match first.
Privacy by architecture
Other chat tools route every visitor message through their cloud and the model vendor's cloud. Zupport.chat loads the AI model into the visitor's browser once, caches it forever, and runs every retrieval and every inference locally.
- No backend inference: there is no server to leak from.
- No conversation logs on our side.
- GDPR, CCPA, HIPAA-adjacent handled by architecture, not policy.
One-line embed, anywhere a script tag lives
Copy one script tag, pin it to a version, paste it into your site. Works on Shopify, WordPress, Webflow, plain HTML, custom React/Vue/Svelte apps — anywhere the browser can run a script.
- Version pinning so updates never break you.
- Themeable via CSS custom properties.
- JavaScript API for open/close/event hooks.
- Framework docs for the popular setups.
<script src="https://github.com/pavelsima/zupport-embed/releases/latest/download/embed.js" defer ></script> <zupport-chat data-assistant-id="your-id"></zupport-chat>
A dashboard you can actually use
Manage assistants, upload sources, curate scenarios, tweak embed configuration, preview live — all in one place. Built on Firebase, so changes you make stream in real time across tabs.
- Sources tab: uploads, crawls, manual subpages, realtime status.
- Scenarios tab: curate, suggest, draft, organize.
- Configure tab: branding, behavior, match threshold.
- Embed tab: copy-ready snippet per assistant.
Works on every device your visitors use
The scenarios layer works everywhere — including iOS, where Safari's WebGPU restrictions prevent on-device LLMs. AI fallback runs on devices that support it, with a graceful consent prompt on Android.
- iOS: scenarios layer always available.
- Android: full two-tier with AI fallback (with consent).
- Desktop: WebGPU-accelerated AI, full speed.
- Graceful degradation on older browsers — scenarios still work.
| Scenarios | AI fallback | |
|---|---|---|
| Desktop | ● | ● |
| Android | ● | ● with consent |
| iOS | ● | — |
Add it to your site today.
Founder pricing — $79 once. First 100 customers only.