Where should we start when adding an AI chat assistant to a website or mini program?
Do not begin by choosing a model or drawing a chat box. Begin with real questions, then define whom the assistant serves, which sources it may rely on, and what happens when it cannot answer. Channel and architecture decisions come afterwards.
If the only requirement is “we want AI customer service,” postpone the API integration. Collect a representative question set from support conversations, site search, sales enquiries, and service tickets. For each question, record the approved answer, source, user identity, whether the information is public, and the consequence of a wrong response. With no history, ask sales, support, and product teams for their most common questions, then test those assumptions with real users.
When defining model, data, and production boundaries, also compare What should you prepare before building an enterprise AI assistant? and Why does an application using a large-model API still answer inaccurately or invent facts?; the linked guidance adds context that should be considered in the same decision.
| First-release scope | Typical work | Required inputs | Recommendation |
|---|---|---|---|
| Navigation and fixed FAQ | Find a page, hours, stable policies | Structured answers or site search | Rules or SaaS may be enough |
| Company knowledge | Product, service, policy, support questions | Versioned, permissioned documents | Retrieve sources and cite them |
| Lead capture | Identify intent and create a contact | Form, CRM, consent record | Explain purpose and minimize fields |
| Transactions | Orders, appointments, ticket changes | Authentication, business API, approval | Add only after Q&A is stable |
A website is well suited to public information, search traffic, and access across channels, and can update independently of mini-program release review. A mini program fits an existing WeChat audience, QR entry, and identity-linked services, but brings platform review, privacy disclosure, and API constraints. Do not launch both automatically. Start on the web when most users arrive there; validate a mini program when WeChat identity or an offline QR journey is essential. The channels can share knowledge, evaluation, and conversation services without having identical interfaces.
For the first release, curate one narrow knowledge package—for example, pre-sales product questions. Give every item a title, scope, publication and expiry dates, owner, and access level. Remove obsolete duplicates and ask the business owner to resolve conflicting statements. The original RAG paper shows how generation can use retrieved external knowledge; retrieval remains only a prerequisite, and results still need evaluation against company questions.
Set explicit conversation rules. If no approved source supports an answer, the assistant says so. Prices, timelines, and policies show their source and effective date. A request for a person triggers immediate handoff. Internal material is never exposed to an unauthenticated visitor. Before collecting a phone number or order ID, explain its purpose and request only what the task needs; the business must approve storage, vendor transfer, retention, and deletion under the applicable privacy rules.
Run three gates before launch. First, evaluate evidence-supported accuracy, severe errors, correct refusal, and citation consistency on held-out real questions. Second, test mobile weak networks, follow-up turns, retries, keyboard overlap, long responses, and handoff; dynamic states such as “answering” and “transferred” should be available to assistive technology, following the W3C WCAG guidance on status messages. Third, confirm that content owners can update sources, support can take over with context, and operators can trace retrieval and model logs.
After launch, track coverage, correct resolution, handoff, unanswered topics, severe errors, P95 latency, and per-conversation cost—not raw chat volume. Wavesteam's responsible first delivery is a question set, source register, answer boundaries, prototype, evaluation report, data flow, and operating console. A chat bubble alone is not the product.