How answers work

Each article is split into passages of about 160 words, keeping Markdown headings with their text. When a visitor asks a question, the assistant ranks passages by keyword relevance (the BM25 method), using English word stems so “refunds”, “refunded” and “refunding” all match “refund”. Passages where the question's words appear close together, or as the same phrase, rank higher.

Confidence check. If the best passages don't contain enough of the question's important words, the assistant doesn't answer. It says it couldn't find an answer and offers “Talk to a human”. The question is listed under Analytics → Unanswered.

With your AI key, the top passages and the question are sent to your AI provider with instructions to answer only from those passages and to cite them as [1], [2]. Answers without valid citations, or where the model says the passages don't cover it, are treated as unanswered.

Without a key, the assistant replies with the most relevant passages and links to their articles. If the AI provider fails, it falls back to the same behaviour.

Synonyms. Common support wording is matched for you: “money back” finds your refund article, “sign in” finds “log in”, “unsubscribe” finds “cancel”, “passcode” finds “password”, “delivery” finds “shipping”, “receipt” finds “invoice”, and so on. Add your own product's terms under Knowledge base → Synonyms, one group per line (for example plan, tier, package). A synonym match counts a little less than the exact word, so the article that uses the customer's own words still ranks first. Retrieval is still keyword-based: it does not understand meaning beyond these variants, so if customers ask in very different words, add those words or a synonym group.