Background
Visitors to a corporate site ask a narrow set of questions: can you build this, roughly what does it cost, how long does it take. The usual answer is to make them hunt across pages or fill in a form and wait. We wanted them to ask on the spot and get an answer that lands on a specific service page rather than a generic pitch.
What we did
- Added a persistent assistant panel in the corner of the site that does not interrupt browsing
- Preloaded six quick questions, one per practice, so visitors do not have to invent a question
- Streamed answers from the server so text appears as it is generated
- Ended each answer with links to the matching service page and to contact details
- Planned a bilingual knowledge source drawn from the Chinese and English service and case content
Technical approach
| Layer | Choice |
|---|---|
| Front end | Vanilla JS panel shipped with the static site |
| Server | Node.js + Express |
| Model access | OpenAI-compatible API, model swappable |
| Transport | SSE streaming (text/event-stream) |
| Credentials | Server-side environment variables only, never in client code |
Results
| Item | Detail |
|---|---|
| Quick questions | 6, one per practice |
| Interface languages | Chinese / English |
| Answer format | Streamed, with service-page and contact links |
| Bilingual knowledge source | Planned |
Authorization
Own product; implementation and architecture are public.