I am trying to create an AI-chatbot with the help of the new beta tool. I am refering to this information page. The goal of the chatbot is helping customers with questions, and sometimes refering them to an employee (when their account is blocked for example). I am using knowledge-base articles to feed content to the AI-bot. The answers the AI-bot is giving varies from really poor to fine (but mostly poor). I dont see much options to train the chatbot expect for making more knowledge articles. What is the best way to train the chatbot? My company currently only uses the knowledge base to feed the AI-chatbot and nothing else. Could there be some problems with using Dutch?
It appears the only way to train the chatbot is via knowledge base articles and public URLs. It doesn't seem like there's a way to force some instructions, at this time. Since it's in beta and still early in the development process, hopefully the capability to control some of the responses will be developed.
I notice this blurb at the bottom of the product update article which makes it seem like English might work the best.
Hey @MeesvP, thank you for posting in our Community!
Creating an effective AI-chatbot requires high-quality training data and proper configuration. To our top experts, @matthew-aire, and @RSchweighartdo you have any recommendations for @MeesvPmatter?
May 23, 20244:59 AM - edited May 23, 20244:59 AM
Member
AI chatbot struggles
SOLVE
Hi @PamCotton , thank you for replying! I am very interested in the suggestions of the experts!
To give another example: When I ask the AI-bot: ''How do I stop my subcription?'', it gives a correct answer. However, when I ask the AI-bot the same question in another session it just replies with ''I dont know'', which is no useful information for customers. It would be nice to receive some tips.
I have exactly the same issue - possibly with Dutch too. I was thinking about creating additional explanations at the end of my knowledge base articles that frame the concepts differently so the bot might understand the problems beter. But would be nice if it would start being more reliable though as it's just using the GPT model, and that has no problem being an agent with the same data
It appears the only way to train the chatbot is via knowledge base articles and public URLs. It doesn't seem like there's a way to force some instructions, at this time. Since it's in beta and still early in the development process, hopefully the capability to control some of the responses will be developed.
I notice this blurb at the bottom of the product update article which makes it seem like English might work the best.