Load message history to core chat from schema chat

[blakerosenthal]

Basic prereq for passing chat history down the pipeline. History is already present in the schema.Chat object loaded from the database, so we just need to make sure it also gets loaded into the ragna.core.Chat object inside schema_to_core_chat.

ragna/ragna/deploy/_api/core.py

Lines 290 to 297 in 647a61e

	with get_session() as session:
	chat = database.get_chat(session, user=user, id=id)
	chat.messages.append(
	schemas.Message(content=prompt, role=ragna.core.MessageRole.USER)
	)
	core_chat = schema_to_core_chat(session, user=user, chat=chat)
	core_answer = await core_chat.answer(prompt, stream=stream)

[blakerosenthal]

@pmeier Tell me if this makes sense:
For the purpose of chat history, we don't need Sources for each individual message since they already exist on the chat object. So we don't need to rebuild them here (?)

[pmeier]

since they already exist on the chat object

They don't. Source's are returned by the SourceStorage and are never saved on the Chat but only ever on the Message

ragna/ragna/core/_rag.py

Lines 221 to 227 in 8ea62a3

	sources = await self._run(self.source_storage.retrieve, self.documents, prompt)
	answer = Message(
	content=self._run_gen(self.assistant.answer, prompt, sources),
	role=MessageRole.ASSISTANT,
	sources=sources,
	)

If you have a look at the ORM model for the chat, we also have no connection to the sources table, but only the messages

ragna/ragna/deploy/_api/orm.py

Lines 75 to 92 in 8ea62a3

	class Chat(Base):
	__tablename__ = "chats"
	id = Column(types.Uuid, primary_key=True) # type: ignore[attr-defined]
	user_id = Column(ForeignKey("users.id"))
	name = Column(types.String, nullable=False)
	documents = relationship(
	"Document",
	secondary=document_chat_association_table,
	back_populates="chats",
	)
	source_storage = Column(types.String, nullable=False)
	assistant = Column(types.String, nullable=False)
	params = Column(Json, nullable=False)
	messages = relationship(
	"Message", cascade="all, delete", order_by="Message.timestamp"
	)
	prepared = Column(types.Boolean, nullable=False)

The messages table however is connected to the sources table

ragna/ragna/deploy/_api/orm.py

Lines 95 to 100 in 8ea62a3

	source_message_association_table = Table(
	"source_message_association_table",
	Base.metadata,
	Column("source_id", ForeignKey("sources.id"), primary_key=True),
	Column("message_id", ForeignKey("messages.id"), primary_key=True),
	)

[pmeier]

LGTM after 8a984a6 is reverted as explained in #429 (comment).

[blakerosenthal]

@pmeier Mypy points out that you can't construct a ragna.core.Document from a schema.Document because it's missing the metadata field. I see a couple ways around that:

1. Add a metadata field to schema.Document, although it seems like that might interact weirdly with the multi-document code that's strewn around that keeps the documents and metadata separate. It would create some redundancies to clean up, but I think it could be nice to just have metadata on the schema object.
2. Cascade all the object conversions directly in schema_to_core_chat kind of like the functions found in database.py.

I'm leaning toward option 1 unless there's a simpler way.

[pmeier]

Ah yeah, I'm hitting the same on a another branch as well. I agree we should go with option 1 especially in the light of #256.

No need to duplicate work here. Just set metadata={} and add a comment that this will be resolved later.