Elasticsearch Document store - embedding retrieval

[anakin87]

Part of deepset-ai/haystack#5329

• Add basic support for Embedding retrieval in Elasticsearch Document Store

Notes for the reviewer

• ElasticSearchEmbeddingRetriever will be added in a subsequent PR.

• I chose to support only Approximate kNN, which is the suggested approach compared to Exact, brute-force kNN.

• This Document Store is compatible with Elasticsearch>=8.11
  Configuring vector fields in older versions required more manual effort on the part of the user (e.g., explicitly specifying the vector size at index creation).

• I haven't implemented scaling scores in the range [0, 1]. See Elasticsearch Document Store - investigate scaling scores for embedding retrieval #53

• I would like to test also other unhappy paths, such as trying to write documents with embeddings of different sizes.
  Concerning this point, we should rework error handling during indexing (write_documents method).

  Update: I encountered several erroneous DuplicateDocumentError while working on embedding retrieval.
  If we do not want to spend time reworking this part at the moment, I would propose printing a warning with the obtained errors. It seems less misleading to me. WDYT?

[silvanocerza]

Let's not match this error but let's check that a custom error is raised when the received embedding have a different size.

[anakin87]

I tried implementing a custom error handling logic, but the code became too messy.
Also, the Elasticsearch exception is quite informative:
elasticsearch.BadRequestError: BadRequestError(400, 'search_phase_execution_exception', 'failed to create query: the query vector has a different dimension [2] than the index vectors [4]')
So, I would rather not handle this error explicitly.

[silvanocerza]

Nice 👍