Knowledge sources

A project can mix any of these. They are ranked together at retrieval time, so a Q&A pair and a crawled page compete on relevance, not on type.

• Website crawl — Helpo AI reads your sitemap (or robots.txt) to discover pages and shows you the list so you choose what to index. Sites without a sitemap accept URLs entered directly.
• File upload — PDF, plain text and Markdown, up to 10 MB each. PDFs have their text extracted; the others are read as-is.
• Help center articles — an article you publish in Helpo is both the page customers read and a source the AI answers from. Editing it re-indexes it; there is no second copy to drift.
• Q&A pairs — one question, one answer. For the refund window, the plan that actually includes the thing, and everything else that was never written down.
• Store catalog — Shopify and WooCommerce products imported from the storefront's public API. Titles, descriptions, prices and stock become answerable, and the AI can show product cards in the chat.

Each source is split on paragraph boundaries into windows of roughly two thousand characters with a short overlap, then embedded and stored against your project. Headings matter: a page with clear section headings produces chunks that each answer one question, which is what retrieval rewards.

Deselect pages that are true today but will not be next month, and pages that say nothing a customer would ask about. Three kinds cause most bad answers:

• Marketing and landing pages — they describe benefits, not behaviour, and the AI will quote the pitch back to a customer with a problem.
• Changelogs and blog posts — they describe how the product used to work, and old entries outrank the current docs for specific feature names.
• Legal pages — long, rarely relevant, and they crowd the retrieval window when someone asks a vague question.

Crawled pages can re-crawl on a daily or weekly schedule on plans that include auto-sync; every plan can refetch a source by hand, subject to a short cooldown that stops accidental re-embedding spend. Re-indexing is atomic: the new chunks go live before the old ones are dropped, so the widget never sees a source with no content. A re-crawl that returns nothing fails instead of emptying the source.

Every source change can notify your team through an automation rule, so a docs update that changes an answer is visible in Slack the moment it lands.