Fine-tuning the large language model

When you upload your business knowledge and perform the initial model training, it will correctly recognize and respond to around 70% of questions that customers may ask about your data. To improve this, you should spend some time fine-tuning the model. After fine-tuning, 90% or even 95% response quality is not unusual.

Fine-tuning can be done in the "Conversations" part of the interface or from the live chat hints panel.

To use fine-tuning, your project should be on the "Company" or "Enterprise" plan with AI engine enabled in the project settings.

Fine-tuning through the conversation history

When you open the conversation log for any user in your project, you will see small icons next to every message sent by the bot or human agent. These icons can be used to fine-tune the ML model, improving the quality of automated responses.

Fine-tuning the AI model through conversations

Click the green checkbox if you like the response, or red cross if you find it irrelevant or poorly formulated. When you mark response as "bad", a small pop-up will appear, asking you to enter the link to the source of the correct reponse. You can skip this, simply marking the response as irrelevant, or provide the model with more info to use in further conversations.

Fine-tuning from the live chat interface

Fine-tuning icons can also be found in the "AI hints" section of a live chat interface.

Fine-tuning the AI model through live chat hints