More RAG and Summarization Metrics

[bnativi]

• add Q&A and RAG metrics that also do text comparison, i.e. LLM guided metrics that use the groundtruth.
• rework Coherence to be specific to summarization (now SummaryCoherence), as that is how coherence is used in the literature. The name was changed to SummaryCoherence because RAGAS recently introduced a non-summarization coherence metric.
• improve documentation.

Metrics added:

• AnswerCorrectness
  • Uses Prediction.annotation.text and Groundtruth.annotation.text
  • RAGAS's answer correctness score is computed as a weighted sum of an f1 score with an answer similarity score. The answer similarity score is computed as an inner product of embeddings. I implemented AnswerCorrectness as just the f1 score, so that we don't require embeddings or an embedding model from the user.
• ContextPrecision
  • Uses Datum.text, Prediction.annotation.context_list and Groundtruth.annotation.text
  • Evaluates a RAG retrieval mechanism using a ground truth (no prediction required).
• ContextRecall
  • Uses Prediction.annotation.context_list and Groundtruth.annotation.text
  • Evaluates a RAG retrieval mechanism using a ground truth (no prediction required).