Port evaluation from 1.x and extend to LLMs

[julian-risch]

We need to extend evaluation features, in particular for RAG pipelines so that users can answer questions like:

• Is this pipeline good enough?
• What should I focus on for optimization?
• Is pipeline A better than B? (performance, costs, latency)

This includes the following components typically appearing in RAG pipelines:
Retrievers, Rankers, DocumentJoiners
a) labels available => statistical metrics
b) no labels available => model based heuristics / pseudo label generator

Generators
a) labels available = model based (SAS, answer correctness ...)
b) no labels = model based (groundedness score)

Tasks

Beta Give feedback

1. Implement statistical-based evaluation and metrics calculation #6061
   19 of 19
   2.x epic topic:eval type:feature +4
   
2. LLM Evaluation in Haystack #6786
   13 of 14
   2.x P1 epic +3