Currently, the Parent document retriever in LangChain takes in two memory sources in the definition, i.e., a vectorStore and a docStore.

retriever = ParentDocumentRetriever(
    vectorstore=vectorstore,
    docstore=store,
    child_splitter=child_splitter,
)

Given MongoDB being a full-fledged database and a vector search, we can both the stages in a single database and correspondingly use a single query pipeline to make this a much efficient query.

Definition of Done:

(a) Create a brief/1-page API Google design doc for the proposed feature that we can put to review with the product and langchain stakeholders
(b) Implement code based on the lgtm'ed design doc