Community: Add sqldocstore support for ParentChildRetriever

[hslee16]

Description

This PR aims to add support for sqldocstore to be used by ParentDocumentRetriever class

ParentDocumentRetriever class is passing in Documents to mset here:
https://github.com/langchain-ai/langchain/blob/master/libs/langchain/langchain/retrievers/parent_document_retriever.py#L135

Issue

Using langchain_community.storage.sql.SQLStore with the ParentChildRetriever causes the following exception:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[7], line 1
----> 1 retriever.add_documents(data, ids=None)

File langchain/retrievers/parent_document_retriever.py:135, in ParentDocumentRetriever.add_documents(self, documents, ids, add_to_docstore, **kwargs)
    133 self.vectorstore.add_documents(docs, **kwargs)
    134 if add_to_docstore:
--> 135     self.docstore.mset(full_docs)

File langchain_community/storage/sql.py:236, in SQLStore.mset(self, key_value_pairs)
    228 self._mdelete(list(values.keys()), session)
    229 session.add_all(
    230     [
    231         LangchainKeyValueStores(namespace=self.namespace, key=k, value=v)
   (...)
    234     ]
    235 )
--> 236 session.commit()

File sqlalchemy/orm/session.py:2032, in Session.commit(self)
   2029 if trans is None:
   2030     trans = self._autobegin_t()
-> 2032 trans.commit(_to_root=True)

File <string>:2, in commit(self, _to_root)
...
...
...

TypeError: can't escape Document to binary

Add tests and docs

A new test has been added to make sure that SqlDocstore supports langchain document types.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Skipped Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)	Visit Preview		Mar 20, 2025 3:33pm

[OskarStark]

I propose to rename the PR title to:

feat(community): add `sqldocstore` support in `ParentChildRetriever`

[a-holm]

Found two potential issues.

Metadata Loss: The _bytes_or_document method currently only stores the Document.page_content and discards the metadata. ParentDocumentRetriever relies on retrieving the full parent document (including metadata) from the docstore. Storing only the content will likely break the retriever's functionality later on, even though the initial mset call succeeds. A better approach would be to serialize the entire Document object (e.g., using pickle or JSON) into bytes for storage, ensuring that mget can retrieve the bytes and potentially deserialize them back into a full Document (perhaps via a wrapper class like EncoderBackedStore or similar patterns used elsewhere).

Inconsistent Type Hint: The type hint for key_value_pairs in mset was updated to include Document, but the corresponding hint in amset was not.

[hslee16]

@a-holm thanks for the comments! I'll address them ASAP.