A Knowledge Base Your Assistant Is Trained On

Nothing is trained: your content goes into a knowledge base from which the answer is searched at runtime. Four source kinds are available — free text, file, website and maintained question-and-answer pairs. A website run reads at most 25 pages, obeys robots.txt, takes only HTML and plain text and follows no link to an external address; pages marked noindex count as a draft and yield no sections at all. Files up to 10240 KB are split into sections, a PDF via its existing text layer. Which source feeds in and when it is read again is your decision per entry. More under knowledge base.

For each answer the conversation run fetches five hits from your knowledge base and builds from them the excerpts the assistant may rely on. If none of the hits carries the question, it is instructed not to guess but to say that the deposited knowledge base holds nothing on the subject. In parallel the question is recorded as a knowledge gap; from the Pro package the list of knowledge gaps shows up to 50 cases and is the working stock for your next round of maintenance. On request the assistant hands over to a human: state change, mailbox entry and notification happen in one go, with the transcript attached. An open question therefore stays visible instead of disappearing into the conversation. More at the support assistant.

The room for that is drawn tight. For every question at most five search hits from your knowledge base, plus your text blocks, go to the model, numbered, and sourced answers name the excerpts they rely on. Figures, prices and deadlines that are not in your sources are replaced with a notice. If the search finds nothing or the model declines, the assistant says that nothing on this is held in the knowledge base — and the case is logged as a knowledge gap, as is an answer without a used excerpt. Its rules forbid it to invent contact details, file numbers or appointments, or to pose as a human being. An error cannot be ruled out entirely with any language model; hence the binding to sources, the visible citation and the handover to a person.

Every source has its own refresh interval: the house interval is 7 days, adjustable between 1 and 365, and the value 0 explicitly means never. A daily run queues the due sources and reads them again; the fetch is conditional via ETag and Last-Modified, so an unchanged page costs no new build. A source counts as stale only once more than twice its interval has passed since the last check, and at least 14 days. Every index run writes into a new version and switches over only at the end, with five versions kept. On request we take over the upkeep as part of ongoing operation. More under knowledge base.

Yes, as long as the document has a text layer. The knowledge base accepts files up to 10240 KB and splits them into overlapping sections so a passage does not end mid-sentence; a PDF is read through its existing text layer. A scanned sheet without a text layer deliberately yields nothing — there is no text recognition from the image, and the reason is shown as a message on the source instead of silently missing. For the question, at most five search hits plus your text blocks then go to the model, numbered, and sourced answers name the excerpts they rely on. For prices, terms or deadlines that have to be exact, the maintained question-and-answer pair is the better place: it goes out verbatim. More under knowledge base.

Ongoing operation starts in the Start package from 49 € per month net including hosting in Germany, care and updates, and the one-off setup with individual training from 990 € net. Start holds ten knowledge sources and 200 documents, Pro from 2,900 € net setup and 149 € per month net holds 50 sources and 2,000 documents, Custom from 4,900 € net setup holds 250 sources and 20,000 documents. A source can be text, a file, a website or a question-and-answer collection. Sourced answers name their source; an answer without a used excerpt is recorded as a gap. Details under knowledge base and pricing. All services can be cancelled monthly; there is no minimum term.

Under Knowledge in your account, per source with a timestamp. It also states whether a crawl was capped and how many links were left open — without that figure the source would afterwards read as finished, as if it had been read in full. As long as nothing has been measured the field stays empty and is not reported as zero. Every source also carries its refresh interval; it counts as stale only once more than twice that interval has passed since the last check, and at least 14 days. A conditional fetch via ETag and Last-Modified may well conclude that nothing has changed. All of this is independent of product changes and the most frequent explanation for changed answers.

Yes. One website is enough for the first answers. The crawler follows the rules in your robots.txt and takes in a limited number of pages per run; from these come the sections the assistant cites in its answers. You can add more at any time: there are four source types — text, file, website and question-and-answer collection. The assistant does not have to be rebuilt for that: every index run writes a new version and only switches over at the end, and a failed run leaves the previous state answering instead of creating a gap.

You do, if you want to: the interface shows sources, documents and conversations, and a changed source is read in again. Every source has its own refresh interval — preset to 7 days, adjustable between 1 and 365, and the value 0 explicitly means: not automatically. A run in the early morning works through the sources that are due; a source only counts as stale once more than twice its interval has passed since the last check, and at least 14 days. Unchanged pages are skipped, and a run that wrote nothing deletes nothing — otherwise a maintenance page would clear out your knowledge base. The analytics show which questions come up often and where an answer is missing. If you would rather not do it yourself, you hand it to us.