It doesn't need OpenAI APIs or anything else external. The system is ready and working. On our website, the chat button is now placed in the lower right corner and provides useful information in real time. I have placed the old chat without AI in the lower left corner as a fallback, which runs via Telegram (re:plain) - even though InsertChat also allows contact via email form (ticket-like) or can move to the side if the live chat is supported (like tawk, sadly re:plain isn't yet). Training is relatively easy. You can upload the sitemap (didn't work right away) or just a list of URLs from the xml, which worked immediately. Furthermore, pdf files (took relatively long, interestingly enough, and some he didn't want, after clicking on fix errors, but then he trained them after all), docx (that always worked very fast; if you have the Word original, so You can also view the activity in the inbox and correct the answers if they were incorrect (brain with plus). I now use GPT 4o mini, which is also recommended by the developers. It is very fast and usually delivers even better results than larger language models like 4o itself. But I like that there are many options for the LLM. If you give the bot too much freedom, the error rate increases. That's why I limited it to the material it has learned and set the creativity to 0.4. This usually gives it reliable and accurate answers.
All in all, it's a good investment, as it greatly speeds up the support process for website visitors and most support requests take care of themselves. The design could really be improved, as others have requested. I would re-design the black borders, mak the less round, nearly square, and only a light gray color. This would look more modern. But functionality is more important, sure :) Since I only train in a specific language (German), it would be nice if I could set it as the default language in the bot settings. I finally took Tier 5. Webhooks are interesting. We don't use Zapier but Make.com.