I’ve been looking at the various use-case examples ppl have been coming up with
Current we have fine tuning, embeddings and edits.
A hot topic is preconfiguring gpt with additional contextual information as to act as a teacher, guide, or whatever…also known as prompt engineering
I suggest adding a endpoint to be able to store these “prompts” server-side
Like this we could send a completion with some input text, and reference a stored prompt configuration
This would prevent us from having to send the full preconfiguration plus input for every request and maxing out on max tokens …and save some bandwidth
Example:
Preconfiguration (xyz) :
Pretend you are a preschool teacher, and explain the following question so that a child can understand it.
Input:
What is quantum computing
Api request:
{
configuration : “xyz”,
prompt : “What is quantum computing”
}
Or even better
Preconfiguration (xyz) :
Pretend you are a {0}, and explain the following question so that a {1} can understand it.
Input:
What is quantum computing
Api request 1:
{
configuration : “xyz”,
interpolations : [“preschool teacher”, “child”],
prompt : “What is quantum computing”
}
Api request 2:
{
configuration : “xyz”,
interpolations : [“quantum physics professor”, “2nd year quantum physics student”],
prompt : “What is quantum computing”
}
These are very simplified examples.
I’ve seen configurations span multiple paragraphs with instructions ranging from language output, to output formatting style, and much more.