Giving the chatbot its knowledge

The chatbot answers from the knowledge you give it, so this is the part that matters most. You manage it on the Chatbot page of the dashboard. For the big picture, start with AI chatbot.

Ways to add knowledge

• Paste text or an FAQ straight into a source.
• Upload a file: .txt, .md, .html or .pdf. PDFs must contain real text; a scanned PDF is refused with a plain explanation.
• Import a web page, or a whole site: the server reads public pages, keeps headings and lists, and drops navigation and scripts. A site import follows links on the same host only, reads at most 12 pages, and honours robots.txt.
• Start from a template: webshop, clinic, restaurant or software. You answer plain-language questions, and only the questions you answer become knowledge.

Keeping it fresh

• Refresh an imported page whenever it changes. A refresh that finds nothing readable keeps the old text.
• Weekly auto-refresh can be switched on per page. Unchanged pages are left alone.
• The last 5 versions of a page are kept, and you can roll back to one. A roll-back also stops the weekly refresh, so it cannot undo itself.

How the bot uses it

• A small base, FAQ-sized, is handed to the model whole, and the model decides which part is relevant. It must hand over when the text does not hold the answer.
• A large base is searched first, and only the best-matching sections are used. If nothing matches, there is no AI answer.
• A synonyms step retries with other words when the first search finds nothing, so "when do you open" can find "opening hours".
• Semantic search (matching by meaning) exists as an optional extra. It is configured by the operator and is not on for everyone.

Testing before you trust it

• Test console: ask a question and see what the bot would do. It is a dry run, so nothing is sent.
• Saved test questions with the outcome you expect, so a change to your knowledge cannot quietly break a good answer.
• Replay: run the last 20 real questions against today's knowledge and see which would now be answered differently.
• Suggestions from your team's answers: when a colleague answers a hand-off, the question and answer are offered as a knowledge entry, with personal details masked. They are offered, never added automatically.
• A weekly summary e-mail lists questions the bot could not answer and knowledge that was never used.

Protecting people's data

• Card numbers, personal ID numbers and IBANs are masked before anything reaches the AI model or a summary. A number split across two messages, or other kinds of secrets, can still slip through, so do not rely on this alone.
• Topics the bot must never discuss: list them, and a message containing one goes straight to a person. This is a safeguard, not a guarantee; other spellings and earlier messages are not caught.

Next: Review, hand-off and trust.