A knowledge base your chatbot answers from

To begin with, what you already have is usually enough. Upload PDFs and Word files, paste in text, or enter your website address. You can pick individual URLs or the whole site. Sources can be combined, for example the manual as a PDF and the FAQ page as a URL.

How many URLs and how much storage an assistant gets depends on your plan. On the free plan, files can be up to 5 MB; higher plans allow larger files.

Some answers should always be worded the same way. Your cancellation terms, a note that the chat does not give legal advice, the question that comes in three times a day. For those you create a Q&A pair and write the answer just as you would give it on the phone.

Q&A pairs are also the quickest fix when the bot gets something wrong. Add the correct answer and check in the Playground that it now comes through. Make sure a pair doesn't contradict your website, otherwise the bot has no way of knowing which statement is right.

Websites change. So the bot doesn't quote last quarter's prices, you can have website sources re-read on a schedule: daily, weekly or monthly. Only content that has changed is processed again, and the assistant is retrained automatically afterwards.

Uploaded files are yours to replace. When there is a new price list as a PDF, swap out the old one and start a new training run. That usually takes a few minutes, and up to 15 minutes for large amounts of data.

The method is called RAG, retrieval-augmented generation. For each question, the assistant first finds the relevant passages in your sources and then phrases its answer, in the language the question was asked in. Your documents can all be in German and a customer from Lyon still gets her answer in French.

Each assistant has its own private vector database. Your content is not shared with other users and is not used to train AI models. Hosting is in Germany or the EU.