Description
Introduction
Since the custom model throws user instructions directly to the AI model. There are two main approaches to sending these instructions,
-
System object: This approach works by sending instructions to the system object which costs fewer tokens but as the context grows it becomes hard for the AI model to remember all the instructions.
-
Few Shot: To achieve more accurate results we are sending user instructions with each message which gives us a lot more accuracy but also costs more tokens since we are sending instructions every time.
Default Behavior
By default, we are using system object approach with each custom model. In case, a user wants to enable few shot approach, the user has to explicitly enable this feature by passing dangerouslyAllowFewShotApproach flag like so:
[
{ ...customModelConfig, dangerouslyAllowFewShotApproach: true }
]