From a folder of PDFs to a bot that answers.

Each document is split into passages and indexed. That happens once, in the background, and you can watch it progress — a file that fails to read says so rather than sitting silently at nought per cent.

Change a price list and re-upload it; the old passages go and the new ones take their place. Nothing you upload is used to train anybody’s model.

The question is matched against your passages, and the most relevant few are given to the model along with it. The answer comes back grounded in your text — which is why it can tell a visitor your Tuesday pallet run goes to Norway, and why it cannot tell them something you never wrote down.

It replies in the language the visitor wrote in. That is not a setting; it is what the model does when the instruction says to.

When your documents do not cover a question, the bot says so and offers to put the visitor through, rather than producing a confident paragraph that sounds right. You write the wording of that fallback, so it sounds like your business.

Every answer is stored with the conversation, so when something does read oddly you can open it and see what the bot was given to work with.

If somebody on your team is online, the visitor waits a moment and is joined in the chat. If nobody picks up within the time you set, the bot collects a name and an email and opens a ticket instead, so the conversation is not lost because it was half past six.

Your team replies from the inbox or straight from their email. The visitor gets a link that reopens the same conversation where they left it.