Implement Sentence-Window Retrieval for benchmark evaluation

[davidsbatista]

The sentence-window approach breaks down documents into smaller chunks (sentences) and indexes them separately.

During retrieval, we retrieve the sentences that are most relevant to the query via similarity search and replace the
sentence with the full surrounding context, using a static sentence-window around the context.

[anakin87]

Might be related: #7389

[julian-risch]

@davidsbatista Could you please clarify in the issue description and title whether this issue includes adding sentence retrieval to our benchmarks? Or is this issue limited to adding a new feature/component to Haystack? I would expect the latter to keep the issue small.

[davidsbatista]

Initially the idea was to have it benchmarked against other architectures over some dataset and I did an initial naive version:

• https://github.com/deepset-ai/haystack-evaluation/blob/sentence-window-retrieval/evaluations/architectures/sentence_window_retrieval.py
• https://github.com/deepset-ai/haystack-evaluation/blob/sentence-window-retrieval/evaluations/evaluation_sentence_window_retrieval.py

But then talking with Mathis and Stefano pointing out the issue, I've been working on extending the DocumentSplitter in a way that implementing sentence-window-retrieval becomes easier.

So the idea for this issue for now is just to have the DocumentSplitter provide the functionality as described in #7389 - then latter we can add an implementation to the haystack-evaluation repo.