Restructure update embeddings

[bogdankostic]

This PR moves parts of the update_embedding method in ElasticsearchDocumentStore and FAISSDocumentStore to the embed_passages method in dense retrievers.

embed_passages now takes as argument a list of Documents instead of a list of strings (=documents) and another optional list of strings (=titles). Furthermore, DensePassageRetriever now contains the instance variable embed_title which can be set to include titles in embedding of passages.

Old Way

dpr = DensePassageRetriever(document_store, embedding_model)

list_of_documents = document_store.get_all_documents()
list_of_strings = [doc.text for doc in list_of_documents]
list_of_titles = [doc.meta["name"] for doc in list_of_documents]

embeddings = dpr.embed_passages(list_of_strings, list_of_titles)

New Way

dpr = DensePassageRetriever(document_store, embedding_model, embed_title=True)

list_of_documents = document_store.get_all_documents()

embeddings = dpr.embed_passages(list_of_documents)

[bogdankostic]

Tests don't work yet with FAISSDocument. There seem to be two problems:

• It's not possible to add documents without meta field
• get_all_documents returns docs without embeddings

[Timoeller]

About the problems

It's not possible to add documents without meta field

@tanaysoni will look into it

get_all_documents returns docs without embeddings

that is unfortunately a feature, not a bug : ) in faiss you cannot get the actual embedding back, but just query and get vector ids. So we have to exclude the embedding value assertion for this case.

[Timoeller]

It's not possible to add documents without meta field

should be fixed with #310

[tholor]

@bogdankostic Please document all breaking changes in your first comment of this PR with a small example old vs new (e.g args for embed_passages). Also add a label "breaking change". Thx!

[Timoeller]

Thanks for adding the detailed notes.

Looking good, lets merge.