feat: Simplification of Document Search Evaluation interface

[azygadlo]

Feature description

Currently, the entry barrier to running a retrieval evaluation example is not so low.

One needs to understand the concepts and initialize objects such as:

• dataloader
• metrics
• pipeline
• evaluator

and then run evaluator.compute() and pass the other 3 objects.

The same functionality could be achieved with a simpler interface, e.g. by initializing only a single object (let's say: EvaluationPipeline), which would handle the data loading and metrics under the hood, in addition to what the current pipeline implementation does.

Furthemore, some reasonable default values can be either set (without having to pass them via a hierarchy of YAML files) or automatically inferred (e.g. based on the existence of OPENAI_API_KEY or other environment variables).

Motivation

The entry barrier should be lower for new users only willing to run example evaluation scripts.

Additional context

initial idea from @mhordynski :

from ragbits.evaluate.recipes.document_search import DocumentSearchEval

eval_recipe = DocumentSearchEval("path/to/my/dataset") \
    .embeddings(
        LiteLLMEmbeddings("text-embeddings-large"),
        LiteLLMEmbeddings("text-embeddings-small"),
        LocalEmbeddings("some-shit-from-hf")
    ).rerankers(
        Reranker1(),
        Reranker2()
    ).rephrasers(
        Reprasher1(),
        Reprasher2()
    )

eval_recipe.evaluate(...)
eval_recipe.optimize(...)

It should wrap lower-level concepts like Evaluator or Optimizer in a higher-level API for simplicity.