Storing and retrieving chat history in a vector db.

[JacobGoldenArt]

Hi, I've been reading through the user guide and specifically looking at the VectorStoreFlowStep. This Retrieval Augmented Generation based on the user prompt seems like it could work well for retrieving both chat_history and possible context that we previously stored in out vector database.

Just thinking out loud: For example , The user asks about blackholes. The 'ChatHistoryFlowStep and the VectorStoreFlowStep could both query their indexes based on the prompt (in parallel). The ChatHistoryFlowStep might return a chat you had a few months back about black holes and the VectorStoreFlowStep might return some relevant data that we stored earlier in our vector db. Now perhaps there is a third flowstep that combines those returns which we finally pass to the ChatOpenAI flow step where it is used as context. (Then there is the other piece where the chat_history is embedded and stored in it's vector database.)

Is there anything already in place for this that I may not have seen? Perhaps an even simpler way would be to just store chat history and all your external data in the same vector db index and just use VectorStoreFlowStep to do it's thing and return the most similar result regardless if it's chat history or external data?

[stoyan-stoyanov]

Great point @JacobGoldenArt! I think the easiest way to do this is to utilize the get_conversation_string method from the MessageHistory() class and save the string as you would save any other data to your vector database. Then you can use the VectorStoreFlowStep as usual and specify top_k > 1 which will (hypothetically) retrieve relevant documents originating from an old message history as well.

I hope this helps!

Let me know if you try it out and if it is straightforward! I've been thinking about creating such example :) If this is not very user-friendly enough we can also consider extending the framework to make this easier.