feat(prompts): The first argument of add_few_shot should be an input model, not a string

[ludwiktrammer]

Currently add_few_shots takes two strings - an arbitrary user message and an arbitrary assistant response to the message.

Instead, the first argument should be of type InputT (i.e., the should be an instance of the pythantic model used to render the prompts). The user message should be then automatically rendered using the user message defined in the prompt class (MyPrompt.user_message) and the input data given as the argument to add_few_shots. This will ensure that the user prompt given in the few-shot examples and the final user prompt are synchronized.

Additionally, the add_few_shots might be able to take both instances of InputT and raw strings - it would use strings directly but when given instances of InputT it would first compile them to strings.

For example, given this pydantic model and prompt:

class WelcomePromptInput(pydantic.BaseModel):
    name: str

class WelcomePrompt(Prompt[WelcomingPromptInput, str]):
    system_prompt = "You are an agent that welcomes the user by name"
    user_prompt = "user's name is {{ name }}"

instead of writing this:

prompt = WelcomePrompt(WelcomePromptInput(name="John"))
prompt.add_few_shot("user's name is Marcin", "Welcome Marcin!")
prompt.add_few_shot("user's name is Beatrice", "Hello Beatrice!")
prompt.add_few_shot("user's name is Alice", "Good Morning Alice!")

you would write:

prompt = WelcomePrompt(WelcomePromptInput(name="John"))
prompt.add_few_shot(WelcomePromptInput(name="Marcin"), "Welcome Marcin!")
prompt.add_few_shot(WelcomePromptInput(name="Beatrice"), "Hello Beatrice!")
prompt.add_few_shot("user's name is Alice", "Good Morning Alice!")

(Note that in reality the prompts might be much longer and more complicated and the values my come from external source like a database)