What to upload
The bot answers from what you give it and nothing else. There is no training step — a document is searchable within a minute or two of finishing its upload, and removing one takes its answers with it.
How it actually reads a document
Uploaded files are split into passages of a few hundred words. When a visitor asks something, the passages closest in meaning to the question are retrieved and handed to the model, which answers from them.
Two consequences follow from that, and they explain almost every disappointing answer:
- The bot never sees a whole document at once. It sees the handful of passages that matched. A fact that only makes sense with three chapters of context around it will not survive the trip.
- Matching is on meaning, not keywords. A visitor asking “can I get my money back” will find a passage headed “Refunds”. They will not find one headed “Section 7.4” that never uses either word.
What makes a good source
- Short pages that answer one question each beat one long document, every time. A page called “Delivery times” retrieves cleanly; the same paragraph on page 47 of a handbook competes with 46 other pages.
- Real headings. They travel with the passage and are a large part of what makes it findable.
- The words your customers use, not the words your industry uses. If people say “shipping” and your document says “fulfilment logistics”, write both somewhere.
- Prices, dates and policies stated plainly. The bot will not infer a price from a table it cannot see the headers of.
What to leave out
- Anything confidential. Everything you upload is material the bot may quote to a stranger. There is no per-document permission model — if it is in the knowledge base, it is answerable.
- Scanned images of text. A PDF that is photographs of pages contains no text to read. Run it through OCR first, or the upload succeeds and the bot learns nothing.
- Out-of-date versions. Two documents that disagree produce a bot that disagrees with itself, and the retrieval has no way to know which one you meant.
When the answer is not there
A bot with nothing to answer from says so rather than inventing something, and offers to pass the question to a person. That is the intended behaviour and it is worth reading what happens next — a “we do not have that” that nobody follows up is only marginally better than a wrong answer.
The Chats page is the best source of new material you have. Questions visitors asked that produced a shrug are exactly the pages worth writing next.