Docs: Fix and supplement answer relevancy description

[jokokojote]

The docs for the answer relevancy metric are wrong in that it is claimed that the score ranges between 0 and 1. As it is written in the calculation section an averaged cosine similarity is used to calculate the metric. The cosine similarity can range between -1 and 1. Of course, most of the time the original question and generated questions will be kind of similar in practice, leading to scores between 0 and 1, but this is not mathematically guaranteed and depends on the used embedder and how it stretches the embedding space.

The following example shows this:

from ragas.metrics import AnswerRelevancy
from langchain_core import embeddings

# demo embedder
class MyEmbeddings(embeddings.Embeddings):

    embedding_lookup = { # some example embeddings
        "Where is France and what is it’s capital?": [0.33, 0.33, 0.333],
         "Pig?": [-0.33, -0.33, -0.33],
        "Orange!": [0.1, 0.2, 0.3],
        "Dark 897897?": [-1000, 0, -1000]
    }

    def embed_documents(self, texts):
        return [self.embedding_lookup[text] for text in texts]

    def embed_query(self, text):
        return self.embedding_lookup[text]
    

answer_relevancy = AnswerRelevancy()
answer_relevancy.embeddings =  MyEmbeddings()

question = "Where is France and what is it’s capital?"

generated_questions = [
    "Pig?", 
    "Orange!", 
    "Dark 897897?"
]

r = answer_relevancy.calculate_similarity(question, generated_questions)
# Score can be negative, depending on embedder:
print(r)

[-0.99999087  0.92721016 -0.8177225 ]

I updated the description and added a formula to make it clearer.

[shahules786]

Thanks buddy @jokokojote You are awesome!