How to Train a Chatbot on PDFs and Documents
Not everything your customers ask about is on your website. Product manuals, return policies, price sheets, guest information and internal notes often live in documents. An AI chatbot can answer from those documents with the same numbered sources as your web pages, so a visitor can see which page of which file an answer came from. This guide shows how to prepare your files, what OneType can and cannot read, and how to keep the chat in step when a document changes.
Step 1: Gather the documents your customers' questions depend on
Start from the questions, not from the folder. Look at the questions customers ask by email and chat, and for each one find the document that holds the answer: a manual, a policy, a FAQ sheet, a menu, a spec. Leave out anything you would not want an answer quoted from, such as internal drafts or documents with personal data. A short list of ten good documents beats a hundred outdated ones.
Step 2: Check that the files can be read
OneType reads five types of file: PDF, Word (.docx), Markdown, plain text and CSV. A file can be up to 25 MB on paid plans and 5 MB on Free, and a PDF can have up to 2,000 pages. PowerPoint, Excel (.xlsx), old Word (.doc) and HTML files are not accepted. OneType also does not read text from images, so a scanned PDF or a photo of a document needs text recognition first: it holds pictures of pages, not text, and there is nothing for the chat to learn from. Open your PDF and try to select a sentence; if you cannot, it is a scan.
Step 3: Upload files or paste text
On the Knowledge page, add Files and drop them in or pick them from your computer; each upload shows its own progress. For a short policy, a list of opening hours or anything you would rather type, add a Text source: give it a title and paste up to 200,000 characters. Headings and lists written in Markdown are kept, so answers can cite the right section. Every file is split into passages that follow its headings, so a source points to the right part of a long document. Headings, lists and tables are kept, and PDF page numbers can show in sources. If a file cannot be read, you get a plain reason, such as a scanned, damaged or password-protected file, or a type OneType does not read.
Step 4: Test the answers and add exact wording where needed
Open the test chat and ask the questions from step 1. Check that each answer shows the right file and, for a PDF, the right page. If an answer is missing or wrong, the cause is usually unclear text in the document: fix the document and upload it again. For answers that must never vary, such as prices, warranty terms or return rules, write a saved answer and choose "Reply with this answer exactly". Text that gives the AI orders, or asks visitors for passwords or card numbers, is refused when you save it, and OneType names the line to remove.
Step 5: Keep documents current
A file is not read again on a schedule. When a document changes, upload a new version. Each file or text source counts as one page toward your plan's page total, and its size counts toward your plan's storage, shared by all agents in your workspace; OneType tells you before a file would go over and says how much room is left. Website pages, in contrast, are read again on the schedule your plan sets. Once a week, look at the questions the chat could not answer, which are grouped by meaning with the most asked first, and decide whether the missing answer belongs in a document, on a page or in a saved answer.
Teach your chat from your documents
See what files work: files and text, learn about saved answers, or start free.