Add RAGAS evaluation metrics

[kcortinas]

closes #317

This PR integrates RAGAS open source evaluation framework for RAG pipelines. The new added metrics offer support for both end-to-end and component-wise RAG evaluation:

• End-to-end evaluation: answer correctness, answer semantic similarity.
• Component-wise evaluation: retrieval metrics (context precision, context relevancy, context recall, context entity recall), generation metrics (faithfulness, answer relevancy).

This PR also fixes:

• Experiment name input when calling generate_metrics
• Missing params when calling query_and_eval_single_line

[martinpeck]

This looks good to me.

I would rather @ritesh-modi give you a final approval, but from my point of view this should be approved and merged.

[martinpeck]

What was the reason for the force-push?

[kcortinas]

What was the reason for the force-push?

@martinpeck The branch became outdated so rebased and pushed with --force-with-lease to avoid overwriting

[ritesh-modi]

Can it be done without using LangchainLLMWrapper and LangchainEMbeddingsWrapper?

[kcortinas]

No. Since RAGAS can be used with different LLM models these wrappers help the RAGAS library implement a common interface used by the metrics. We could of course create our own implementation of the wrappers, but it would require duplicating a lot of existing logic and continuous monitoring of the RAGAS library for compatibility issues. Is there a specific reason why we would want not to import them?

[ritesh-modi]

This needs to be evaluated because of new security guidance

[guybartal]

@ritesh-modi , we have many LangChain dependencies already which we need to resolve with priority,
I think we should open a new separate issue to address this across and not block this PR to prerelease branch.
WDYT?

[guybartal]

LGTM, if all comments are resolved, lets merge this to prerelease branch

[kcortinas]

As discussed, closing this PR per security guidance