Context

apoorva.joshi@mongodb.com pointed out that she wasn't seeing any speed improvements in using the ParentDocRetriever. I looked into this and see that that's because we do not provide a function that can take advantage of concurrency in calling out to the LLM to create embedding vectors when we add documents to the Collection.

A batching structure is already in place, so this really is going from for to async for

Note that one can call aadd_documents now, but this is implemented in the super-class as await run_in_executor(self.add_documents, *, ** so it doesn't provide benefit to the calls to the LLM, which can be done via aembed_documents.

So, we implement
_aadd_documents
_aadd_texts
and abulk_embed_and_insert_texts

and call out to
{{aembed_documents(texts)

Definition of done

Implement. Add tests. Open PR. Merge. Release.