Langgraph extension experiment

[helenCNode]

The langgraph branch tested a simple graph for making retrieval more sophisticated ('retrieval graph').

This branch tests a simple graph for commenting on or extending LLM results ('chat graph').

The following changes have been necessary in order to make the code as simple, flexible and comprehensible as possible for future development.

• The State class representing inputs to graph nodes now inherits from LangGraph's MessagesState, which future collaborators will recognise easily from LangGraph documentation and examples
• To enable this, the old Response class representing chain output has become CustomAIMessage, a class which inherits from LangChain's AIMessage
• Within the graph, responses are now converted into CustomAIMessages and appended to state['messages']

The simple graph has a node which comments on whether the LLM response is likely to be using up-to-date information, based on context metadata.

I think future deployments need to avoid this split between retrieval and chat graphs and just have one graph representing the entire workflow.

[beingkk]

Hey @helenCNode just raising the attention that there are conflicts with the dev branch