feat: Add SimilarityRanker to Haystack 2.0

[vblagoje]

Why:

Ranking by similarity to a query is essential. Enter SimilarityRanker.

What:

Added SimilarityRanker to the rankers package. It's all about evaluating document relevance based on similarity.

How can it be used:

Here's a quick dive into its usage:

from haystack.preview import Document
from haystack.preview.components.rankers import SimilarityRanker

ranker = SimilarityRanker()
docs = [Document(text="Sarajevo"), Document(text="Berlin")]
query = "City in Bosnia and Herzegovina"
output = ranker.run(query=query, documents=docs)
docs = output["documents"]

assert len(docs) == 2
assert docs[0].text == "Sarajevo"

How did you test it:

Unit tests are in place. Also, ran the above snippet. Worked as expected.

Notes For Reviewer:

Reviewer, please do a deep dive into the similarity metrics and integration. Let's ensure this fits seamlessly into Haystack's architecture.

[ZanSara]

Just some comments. nothing worrying 😊

[ZanSara]

No need to call super()

[vblagoje]

cc @sjrl

[silvanocerza]

I would just go with model to be fair, type hints and documentation can do the rest.

[sjrl]

Thanks for the work on this! I have a few more comments about potential additional features.

1. The addition of embed_meta_fields as shown here from V1

   haystack/haystack/nodes/ranker/sentence_transformers.py

   Lines 131 to 133 in 6424354

   	docs_with_meta_fields = self._add_meta_fields_to_docs(
   	documents=documents, embed_meta_fields=self.embed_meta_fields
   	)

   
   Sol has found this to be immensely useful just as we have for EmbeddingRetrievers in haystack V1. I think it would be a great addition to add here as well. We actively use this feature in client projects. And you can see how this is implemented for Embedding models in V2 here

   haystack/haystack/preview/components/embedders/sentence_transformers_document_embedder.py

   Line 26 in 6424354

   metadata_fields_to_embed: Optional[List[str]] = None,

2. Being able to optionally specify a top_k here like we do in V1 would also be helpful I think since we will often pipe the output of this node directly into PromptNode. And I think optionally would be good in case we use nodes like the TopPSampler instead.

[vblagoje]

@sjrl let's keep track of these enhancement suggestions (actually existing V1 features :-) ), but for now, I propose we integrate the simplest implementation we currently have so we can experiment and build demos.

[vblagoje]

PR is awaiting #5945 more specifically, Document unfreezing. Current test failures are expected.

[vblagoje]

@sjrl and @silvanocerza would you give it one last pass please?

[sjrl]

Looks great! Only one minor comment about additional test.

[vblagoje]

Should be gtg now, please have another look @sjrl

[sjrl]

I'm glad that helped catch a bug!