Knowledge Sources

Where does the AI get its knowledge?

Everything the AI uses splits into two groups, and keeping them apart makes it much easier to decide where a new piece of information belongs.

• Facts tell the AI what is true: your articles, your website, your product catalogue, your past conversations. It searches these while it works and grounds its reply in what it finds.
• Instructions tell the AI what to do: your policies, your workflows, your tone of voice. These are applied, not searched for and maybe used.

A wrong fact makes an answer inaccurate. A missing instruction makes it unhelpful even when every fact in it is correct. Most topics need both.

Note: when sources contradict each other, policies win. The AI is instructed to follow the policy and disregard anything conflicting from the catalogue, the website or an old conversation. That makes a policy the fastest way to overrule information you cannot fix at the source right away.

Your articles 📄

Articles are the primary source and the one you control most directly. Every published article is chunked, embedded and searched on every conversation, within seconds of you saving it.

• Public articles ground customer-facing answers and are built into your help center as pages.
• Internal articles are never published, but still ground the AI. This is where supplier arrangements, goodwill limits and how far an agent may stretch a rule belong.
• Imported articles from a previous helpdesk count too, once someone has checked they are still accurate.

They live under Knowledge → Articles. See the Articles guide for the fields and for how to write one the AI can actually use.

Your policies

A policy is a short, standing rule: what you will and will not do, what to never promise, when to hand over. It is not one fact among many, it is the decision your business has already made, and it is applied whether or not a search happens.

Rule of thumb: use a policy for something that holds across many conversations, and an article for a fact a customer might ask about. Policies live under Knowledge → Policies.

Your workflows

A workflow is the step-by-step procedure for one situation: cancel an order, handle a damaged parcel, change a booking. The AI picks the workflow that matches, then follows its steps, checks its conditions and calls the tools it needs along the way. Use one when the steps matter more than the wording, and when the answer depends on data that has to be fetched first.

Your website 🌐

We crawl your site and index what is there: shipping and returns pages, terms, guides, category and landing pages, the FAQ that never became an article. It fills the gaps between your written articles and it keeps up with your site as it changes. Crawling is per market, so each storefront is indexed in its own language.

You can add a Website source yourself under Settings → Knowledge Base, and see what actually got indexed under Knowledge → External Content.

Note: a website crawl is a supplement, not a substitute. Site copy is written to sell, so it is usually vaguer than a support answer needs to be. When the AI keeps getting one topic slightly wrong from your website, that topic wants an article.

Your product catalogue 🛒

Products come from a feed rather than a crawl, which is what makes them reliable: names, descriptions, variants, prices, stock, images and links, refreshed on a schedule. That is how the AI recommends a specific item, answers "will this fit?" and links straight to the right page instead of describing something vaguely.

• Google Shopping Feed you can add yourself under Settings → Knowledge Base.
• Shopify product and collection feeds and your own JSON feed are set up with us during onboarding.

Because it is a feed, a price or stock change reaches the AI on the next refresh. Nobody has to re-upload anything.

Documents and PDFs 📎

Plenty of the real answer lives in a PDF: assembly instructions, a datasheet, a care guide, a size chart, a course book. We can pull those in, extract the text and attach it to the products they belong to, so the AI answers from page 14 of the manual instead of guessing. It works when the documents are reachable from your feed, your PIM or your product pages.

Note: there is no self-serve upload box for documents today. If you have a body of PDFs you want the AI to know, talk to us and we will get it indexed. In the meantime the fastest route is often to put the content on a page on your own site, where the crawl picks it up by itself.

Your past conversations 💬

Your resolved tickets are knowledge too. The AI can search how your team has answered a question before and use those replies as a reference. That matters most in the first weeks, when your articles are still thin but years of good answers are already sitting in your history.

It is a reference, not a rule. A past answer that was wrong stays wrong, which is why articles and policies always outrank it.

Live data from your systems 🔌

Order status, tracking, subscriptions and customer records are not knowledge. They are data the AI fetches while it works, through the tools and integrations you have connected, and nothing about them is indexed in advance. That is deliberate: the answer to "where is my parcel?" should be looked up at the moment it is asked, never remembered.

What the model already knows 🌍

Underneath all of this sits a large language model with broad general knowledge of the world, learned from a very large amount of public text. That is what lets it write fluent Danish, German and English, make sense of a badly typed question, know what a customs charge is, and understand that a customer writing "it arrived in pieces" is upset.

What it must never do is answer questions about your business from that general knowledge. Your prices, policies, delivery times and stock are yours, and only your sources can supply them. The AI is instructed to ground business answers in what it retrieved, to link to the article or page behind an answer, and to hand over to a human rather than fill a gap with something that merely sounds plausible.

So when the AI gets something wrong, the cause is almost always a missing, outdated or contradictory source, not the model. That is good news: it means the fix is something you can write.

Good to know

• Unpublishing an article removes it from the AI's index as well, which makes the Published toggle the quickest way to pull wrong information out of circulation while you rewrite it.
• Knowledge is scoped per market, so an article, a crawl and a feed all belong to a language and storefront. An article with no market is shared across all of them.
• Two articles stating different return windows force the AI to pick one. Keep a single source of truth per fact and delete the duplicate.
• Whenever a colleague corrects the AI, the correction belongs in an article or a policy. Fixing only the reply teaches the system nothing.
• A good first pass: connect your website and product feed for broad coverage, write policies for the rules your team knows by heart, then write articles for your top twenty questions, one topic each.