An agent that knows your data

Ask a general-purpose model what your return window is and it will give you a confident, plausible, completely invented answer. It has no way to know. The entire value of a support agent is that it answers from *your* material, and the quality of that material is the single biggest lever on whether the agent is useful or embarrassing.

That is why there are four source types rather than one. Real knowledge is scattered. Some of it is in a PDF, some is on a help page, some was never written down at all, and a small, important core of it has to be word-perfect. Each of those needs a different way in.

Give the crawler a root URL and it will follow every link beneath it. This is usually the fastest way to get an agent from nothing to genuinely useful, because most companies have already written their documentation as a website.

The step that matters is the one people skip. Before you train, the crawler hands back the list of pages it found, and you can edit it. Delete the careers page, the press releases, the four-year-old blog posts. An agent trained on your entire site will cheerfully answer questions about a job opening you closed in 2023. Ten seconds of pruning here is worth more than any amount of prompt tuning later.

Note that the crawler reads pages, not files hosted on those pages. A PDF linked from your docs is not picked up by the crawl. Upload it directly.

Some answers cannot be paraphrased. Pricing, refund windows, warranty terms, anything with a number in it that a customer might hold you to. For these, use Q&A pairs: you write the question and you write the answer, and you can attach several phrasings of the same question to one response.

Everything else is better as a document or a crawl, because the agent can synthesise across them. Use Q&A sparingly and deliberately, for the small set of answers where being approximately right is the same as being wrong.

Files are PDF, DOCX or TXT, up to 42MB each, uploaded one at a time. Your total storage and the number of links you can crawl depend on your plan: the Basic tier allows 50 links per agent and Pro and Business are unlimited. See pricing for the details.

Two things catch people out. First, PDFs need a real text layer. A scanned document is an image as far as the agent is concerned, and it will extract nothing from it. If you can select the text in your PDF reader, you are fine. Second, retraining is something you do, not something that happens. If you update a document or your site changes, the agent does not notice on its own. You retrain the source that changed. It takes one click, but it is a click somebody has to make.

Once the agent knows your material, decide which model should reason over it on the model selection page, and give it the ability to look things up live with agent actions for anything too fast-moving to train on.