import os import uuid import json from datetime import datetime from dotenv import load_dotenv load_dotenv() from supabase import create_client from langchain_core.runnables import RunnableConfig from typing import TypedDict, List, Optional, Annotated, Literal from langchain_core.messages import BaseMessage, AIMessage, HumanMessage, SystemMessage from langgraph.graph.message import add_messages from langgraph.graph import StateGraph, END from langchain_openai import OpenAIEmbeddings from langchain.chat_models import init_chat_model from langchain_pinecone import Pinecone as LangchainPinecone from langchain_core.chat_history import BaseChatMessageHistory from langchain_core.messages import messages_from_dict, messages_to_dict from pydantic import BaseModel, Field from relevant_links import RELEVANT_NATURALIZATION_LINKS from prompts import SYSTEM_MESSAGE_TEXT, ROUTER_SYS # --- Supabase Setup --- SUPABASE_URL = os.getenv("SUPABASE_URL") SUPABASE_KEY = os.getenv("SUPABASE_KEY") supabase = create_client(SUPABASE_URL, SUPABASE_KEY) # --- Custom State --- class CustomState(TypedDict): messages: Annotated[List[BaseMessage], add_messages] user_id: str chat_history: Optional[List[List[str]]] # --- Router Structured Output Schema (SINGLE MODEL) --- class RouterDecision(BaseModel): action: Literal["reply", "retrieve"] language: Literal["de", "en", "tr", "ar"] text: Optional[str] = None normalized_query: Optional[str] = None # --- AI Models --- if not os.environ.get("OPENAI_API_KEY"): os.environ["OPENAI_API_KEY"] = input("Enter API key for OpenAI: ") llm = init_chat_model("gpt-4.1", model_provider="openai", temperature=0.5) # router with structured output using the single model _router_llm_base = init_chat_model("gpt-4.1", model_provider="openai", temperature=0.7) router_llm = _router_llm_base.with_structured_output( schema=RouterDecision, method="json_schema" ) # --- Vector Store --- pinecone_api_key = os.getenv("PINECONE_API_KEY", "").strip() pinecone_index_name = os.getenv("PINECONE_INDEX_NAME", "").strip() embeddings = OpenAIEmbeddings(model="text-embedding-3-large") vector_store = LangchainPinecone( index_name=pinecone_index_name, embedding=embeddings, pinecone_api_key=pinecone_api_key ) # --- Supabase Chat Memory --- class SupabaseChatMessageHistory(BaseChatMessageHistory): def __init__(self, user_id: str, supabase_client): self.user_id = user_id self.supabase = supabase_client @property def messages(self): response = self.supabase.table("chat_memory").select("*").eq("user_id", self.user_id).execute() data = response.data[0] if response.data else {} return messages_from_dict(data.get("messages", [])) def add_message(self, message): all_messages = messages_to_dict(self.messages + [message]) self.supabase.table("chat_memory").upsert({ "user_id": self.user_id, "messages": all_messages }).execute() def add_user_message(self, content: str): self.add_message(HumanMessage(content=content)) def add_ai_message(self, content: str): self.add_message(AIMessage(content=content)) def clear(self): self.supabase.table("chat_memory").delete().eq("user_id", self.user_id).execute() class MemoryStore: def get_memory(self, user_id: str): return SupabaseChatMessageHistory(user_id, supabase) store = MemoryStore() # --- Logging --- def log_interaction_cloud(user_id: str, question: str, answer: str): payload = { "user_id": user_id, "question": question, "answer": answer, "timestamp": datetime.utcnow().isoformat(), } try: res = supabase.table("logs").insert(payload).execute() print("[Supabase Log Success]:", res.data) except Exception as e: print("[Supabase Log Error]:", e) # --- Retrieval --- def retrieve(query: str): retrieved_docs = vector_store.similarity_search(query, k=4) serialized = "\n\n".join( (f"[Source]({doc.metadata.get('source', '')}):\n{doc.page_content}") for doc in retrieved_docs ) return serialized, retrieved_docs def generate_user_id(): return "user_" + uuid.uuid4().hex[:6] # --- STRICT Router (Structured Output + Retry + Runtime guards) --- def route_or_reply(user_text: str, prior_msgs: List[BaseMessage]) -> dict: history_slice = prior_msgs[-4:] if prior_msgs else [] msgs = [SystemMessage(ROUTER_SYS)] + history_slice + [HumanMessage(user_text)] last_err = None for attempt in range(2): try: obj = router_llm.invoke(msgs) data = obj.model_dump() action = data.get("action") if action == "reply": if not data.get("text"): raise ValueError("RouterDecision 'reply' requires 'text'.") data.pop("normalized_query", None) elif action == "retrieve": if not data.get("normalized_query"): raise ValueError("RouterDecision 'retrieve' requires 'normalized_query'.") data.pop("text", None) else: raise ValueError(f"Unknown action: {action}") return data except Exception as e: last_err = e print(f"[Router Parse/Validation Error Attempt {attempt+1}]:", repr(e)) return {"action": "reply", "language": "en", "text": "Sorry, something went wrong. Please try again."} # --- Main Node --- def query_or_respond(state: CustomState) -> CustomState: user_id = state.get("user_id") or generate_user_id() human_messages = [m for m in state["messages"] if m.type == "human"] latest_user_msg = human_messages[-1] memory = store.get_memory(user_id) history = memory.messages decision = route_or_reply(latest_user_msg.content, history) print("[Router]", decision) if decision.get("action") == "reply": response = AIMessage(content=decision.get("text", "")) memory.add_user_message(latest_user_msg.content) memory.add_ai_message(response.content) #log_interaction_cloud(user_id, latest_user_msg.content, response.content) return {**state, "messages": [response], "user_id": user_id} normalized_query = decision.get("normalized_query") or latest_user_msg.content retrieved_text, _ = retrieve(normalized_query) history_tail = history[-5:] system_msg = SystemMessage( SYSTEM_MESSAGE_TEXT + "\n\nStrictly use the following documents to answer the question. " "If helpful, include direct source links in your answer: " + retrieved_text ) prompt = [system_msg] + history_tail + [latest_user_msg] response = llm.invoke(prompt) print("The retrieved text is ", retrieved_text) memory.add_user_message(latest_user_msg.content) memory.add_ai_message(response.content) log_interaction_cloud(user_id, latest_user_msg.content, response.content) return {**state, "messages": [response], "user_id": user_id} # --- Graph --- graph_builder = StateGraph(CustomState) graph_builder.add_node("query_or_respond", query_or_respond) graph_builder.set_entry_point("query_or_respond") graph_builder.add_edge("query_or_respond", END) graph = graph_builder.compile() # --- Gradio UI --- import gradio as gr def clear_history(user_id_state): if user_id_state: store.get_memory(user_id_state).clear() return [["", "Ich bin hier, um Ihnen bei allen Fragen zur Einbürgerung in Deutschland zu helfen..."]], user_id_state, "", user_id_state def respond(user_message, chatbot, user_id_input, chat_state, user_id_state): if not user_message: return chatbot, chat_state, user_id_input, "", user_id_state user_id = user_id_state or user_id_input or generate_user_id() state: CustomState = { "messages": [HumanMessage(content=user_message)], "user_id": user_id, "chat_history": None } config = RunnableConfig(configurable={"session_id": user_id}) final_state = graph.invoke(state, config=config) bot_response = final_state["messages"][-1].content chatbot.append([user_message, bot_response]) return chatbot, chatbot, user_id, "", user_id description = """ ### Was ist **Einbürgerungs-Helfer**? **Deutsch:** Einbürgerungs-Helfer hilft dir, den Weg zur deutschen Staatsangehörigkeit zu verstehen – einfach, mehrsprachig, datenschutzfreundlich. """ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue")) as iface: gr.Markdown(f"# Einbürgerungs-Helfer\n{description}") with gr.Row(): user_id_input = gr.Textbox(label="Benutzer-ID (User ID)", placeholder="z.B. user_123", lines=1) clear_btn = gr.Button("🗑️ Verlauf löschen") user_id_state = gr.State(value="") chat_state = gr.State([]) chatbot = gr.Chatbot( label="Chat", show_copy_button=True, value=[[None, "Ich bin hier, um Ihnen bei allen Fragen zur Einbürgerung in Deutschland zu helfen.\nLassen Sie uns loslegen – wie kann ich Ihnen heute helfen?"]] ) msg_input = gr.Textbox(placeholder="Frage eingeben...", label="Deine Nachricht", lines=1, value="") with gr.Row(): submit_btn = gr.Button("Senden", variant="primary") submit_btn.click(fn=respond, inputs=[msg_input, chatbot, user_id_input, chat_state, user_id_state], outputs=[chatbot, chat_state, user_id_input, msg_input, user_id_state]) msg_input.submit(fn=respond, inputs=[msg_input, chatbot, user_id_input, chat_state, user_id_state], outputs=[chatbot, chat_state, user_id_input, msg_input, user_id_state]) clear_btn.click(fn=clear_history, inputs=[user_id_state], outputs=[chatbot, user_id_input, msg_input, user_id_state]) iface.launch(server_name="0.0.0.0", server_port=int(os.getenv("PORT", 7860)), share=True)