System Instruction
In this section, you will find tips and guides on writing instructions and knowledge base files.
Last updated
In this section, you will find tips and guides on writing instructions and knowledge base files.
Last updated
Let's start with the fact that the quality of Suvvy's work largely depends on the quality of writing and structure of its system prompt (instruction). Since the foundation of Suvvy is LLM models, general rules must be followed when writing it.
A basic instruction is created when the bot is set up:
We suggest structuring the instruction in a way where each section is marked with a # sign and the model can easily identify where a new section starts and ends.
In the comment to the right of the section, it is described what needs to be written under the heading.
In the #GENERAL INFORMATION section, we recommend describing:
information about the company - what it does, in which region it operates, what products it provides - in general terms without going into specifics. Details are then specified in the knowledge base.
information about the bot itself - its role, who it is, what its name is, etc.
general information the bot should know each time it is addressed
In the #GOALS AND OBJECTIVES section, we recommend describing:
Specific recommendations for the bot to follow, such as asking specific questions, providing information to the client, etc.
The more detailed, YET concise you describe this section, the better the bot will perform in answering.
IMPORTANT!
You can create separate sections within the instruction for specific situations that the bot should be aware of, similar to other sections, marking them with the # sign.
The essence of the #GREETING block is to describe how the bot should greet its interlocutor. Here you can describe the greeting text or any individual rules.
In the #RESPONSE LANGUAGE block, you can specify the conditions under which the bot will communicate with the interlocutor.
The #RESPONSE STYLE block addresses everything related to the style of the bot's responses. By changing the instruction in this part, you can achieve radically different responses. Here you can reduce amiability, shorten response length, etc.
