feat: few shot example optimzier

[jjmachan]

optimize with few short examples

from ragas.metrics import AspectCritic
from ragas.llms import llm_factory

# define metric
llm = llm_factory("gpt-4o")
metric = AspectCritic(
    name="answer_correctness",
    definition="Given the user_input, reference and response. Is the response correct compared with the reference",
    llm=llm,
)

# optimize with annotation
from ragas.config import DemonstrationConfig
demonstration_config = DemonstrationConfig()
metric.train(
    "alignment_sample.json",
    demonstration_config=demonstration_config,
)

[shahules786]

In my experiments embedding user_input only and whole input (containing user_input, response,etc) gave almost identical results. But I think logically it makes more sense to only embed user_input. Although this has to be revisited again later. Let us leave it like this for now.

[jjmachan]

but what if there is no user_input?

[shahules786]

It's something we have to figure out, ie what's the most accurate way to embed few shot examples. Let's leave it like this now.

[shahules786]

Overall looks good. Logic and everything is perfect. Suggested few changes.