If your KB lives inside WordPress (a custom post type, regular posts under a support category, or block patterns) SleekAI reads it natively, no export step. For docs hosted elsewhere (Zendesk, Intercom, Notion) the practical path is to mirror the Markdown into an OpenAI Files vector store. Either way the chatbot widget still runs on your WordPress install, where the rest of your support pages already live.

Yes. The system prompt instructs the model to end every reply with a link to the article that informed it. For technical content this is non-negotiable: users will not trust an unsourced claim about backups or upgrade paths. Citations also create a feedback loop, if the link points at the wrong section, you fix the article and the bot improves on the next query.

Stuffing the full KB into every prompt is wasteful and inaccurate. Use OpenAI Files as a vector store, SleekAI retrieves only the articles relevant to the question, keeps token use predictable, and stays accurate even on very large references. Some teams split further with multibot, a 'billing bot' and a 'sync bot' each scoped to a smaller corpus.

The system prompt encourages the model to ask one clarifying question when the symptom is ambiguous (browser, OS, plugin version, error message) before jumping to a fix. That single follow-up usually halves the number of wrong-article suggestions. For longer flows, structure the KB around named diagnostic paths the bot can quote step by step.

Yes. Display conditions scope a chatbot to URL patterns, post types, or logged-in roles, and multibot lets you spin up several focused assistants: 'login bot', 'sync bot', 'billing bot', each with its own scope and system prompt. A visitor on /help/login gets the login bot, a visitor on /help/billing gets the billing one.

SleekAI is a one-time WordPress plugin license, not a per-conversation SaaS subscription. You bring your own OpenAI, Anthropic, Google, or OpenRouter API key, so you pay the model provider directly at standard API rates, usually a fraction of what white-label support bots charge. No token markup, no resolved-conversation fee, no monthly seat cap.

Two levers. First, the system prompt explicitly forbids invented steps and tells the model to defer when context is missing. Second, retrieval keeps the model grounded, if the right article is in context, hallucination drops sharply. Conversation logs let you spot any drift and tighten the prompt or fill the doc gap that caused it.

Yes. The bot can collect the user's email, the symptom, the steps it suggested, and the environment data (browser, plugin version, account ID) and forward the payload to Help Scout, Zendesk, Freshdesk, or your CRM via webhook. The human reply starts with all the context already gathered, instead of asking the same questions a second time.