src/state.py

from langgraph.graph import START, StateGraph
from langgraph.checkpoint.memory import MemorySaver
from langchain_core.messages import HumanMessage, AIMessage, BaseMessage
from typing import Sequence
from typing_extensions import TypedDict, Annotated

class State(TypedDict):
    input: str
    chat_history: Annotated[Sequence[BaseMessage], "add_messages"]
    context: str
    answer: str
    
# For direct RAG chains
def build_graph(rag_chain):
    def call_model(state: State):
        response = rag_chain.invoke(state)
        return {
            "chat_history": [
                HumanMessage(state["input"]),
                AIMessage(response["answer"]),
            ],
            "context": response.get("context", ""),
            "answer": response.get("answer", "")
        }

    workflow = StateGraph(state_schema=State)
    workflow.add_node("model", call_model)
    workflow.add_edge(START, "model")
    #memory = MemorySaver()
    #return workflow.compile(checkpointer=memory)
    return workflow.compile() # Stateless; relies on session memory only


# For agent_chain.invoke
def build_graph_with_callable(call_fn):
    def call_model(state: State):
        response = call_fn({"input": state["input"]})
        return {
            "chat_history": [
                HumanMessage(state["input"]),
                AIMessage(response.get("output", response.get("answer", ""))),
            ],
            "context": response.get("context", ""),
            "answer": response.get("output", response.get("answer", ""))
        }

    workflow = StateGraph(state_schema=State)
    workflow.add_node("model", call_model)
    workflow.add_edge(START, "model")
    return workflow.compile()