import uuid import threading import asyncio import json import re from datetime import datetime from fastapi import FastAPI, WebSocket, WebSocketDisconnect from langchain_core.messages import AIMessage, HumanMessage, SystemMessage from langgraph.graph import StateGraph, START, END import faiss from sentence_transformers import SentenceTransformer import pickle import numpy as np from tools import extract_json_from_response, apply_filters_partial, rule_based_extract, format_property_data, estateKeywords import random from langchain_core.prompts import ChatPromptTemplate from langchain_core.tools import tool from langchain_core.callbacks import StreamingStdOutCallbackHandler, CallbackManager from langchain_core.callbacks.base import BaseCallbackHandler from transformers import AutoTokenizer, AutoModelForCausalLM, TextStreamer class CallbackTextStreamer(TextStreamer): def __init__(self, tokenizer, callbacks, skip_prompt=True, skip_special_tokens=True): super().__init__(tokenizer, skip_prompt=skip_prompt, skip_special_tokens=skip_special_tokens) self.callbacks = callbacks def on_new_token(self, token: str): for callback in self.callbacks: callback.on_llm_new_token(token) class ChatQwen: def __init__(self, temperature=0.3, streaming=False, max_new_tokens=512, callbacks=None): self.temperature = temperature self.streaming = streaming self.max_new_tokens = max_new_tokens self.callbacks = callbacks self.model_name = "Qwen/Qwen2.5-1.5B-Instruct" self.tokenizer = AutoTokenizer.from_pretrained(self.model_name) self.model = AutoModelForCausalLM.from_pretrained( self.model_name, torch_dtype="auto", device_map="auto" ) def generate_text(self, messages: list) -> str: """ Given a list of messages, create a prompt and generate text using the Qwen model. In streaming mode, uses a TextIteratorStreamer and iterates over tokens to call callbacks. """ # Create prompt from messages using the tokenizer's chat template. prompt = self.tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) model_inputs = self.tokenizer([prompt], return_tensors="pt").to(self.model.device) if self.streaming: from transformers import TextIteratorStreamer from threading import Thread # Create the streamer that collects tokens as they are generated. streamer = TextIteratorStreamer(self.tokenizer, skip_prompt=True, skip_special_tokens=True) generation_kwargs = dict( **model_inputs, max_new_tokens=self.max_new_tokens, streamer=streamer, temperature=self.temperature, do_sample=True ) # Run generation in a separate thread so that we can iterate over tokens. thread = Thread(target=self.model.generate, kwargs=generation_kwargs) thread.start() generated_text = "" # Iterate over tokens as they arrive. for token in streamer: generated_text += token # Call each callback with the new token. if self.callbacks: for callback in self.callbacks: callback.on_llm_new_token(token) # In streaming mode you may want to return empty string, # but here we return the full text if needed. return generated_text else: outputs = self.model.generate( **model_inputs, max_new_tokens=self.max_new_tokens, temperature=self.temperature, do_sample=True ) # Remove the prompt tokens from the output. prompt_length = model_inputs.input_ids.shape[-1] generated_ids = outputs[0][prompt_length:] text_output = self.tokenizer.decode(generated_ids, skip_special_tokens=True) return text_output def invoke(self, messages: list, config: dict = None) -> AIMessage: config = config or {} # Use provided callbacks if any, otherwise default to the callbacks in the instance. callbacks = config.get("callbacks", self.callbacks) original_callbacks = self.callbacks self.callbacks = callbacks output_text = self.generate_text(messages) self.callbacks = original_callbacks if self.streaming: return AIMessage(content="") else: return AIMessage(content=output_text) def __call__(self, messages: list) -> AIMessage: return self.invoke(messages) class WebSocketStreamingCallbackHandler(BaseCallbackHandler): def __init__(self, connection_id: str, loop): self.connection_id = connection_id self.loop = loop def on_llm_new_token(self, token: str, **kwargs): asyncio.run_coroutine_threadsafe( manager_socket.send_message(self.connection_id, token), self.loop ) llm = ChatQwen(temperature=0.3, streaming=True, max_new_tokens=512) index = faiss.read_index("./faiss.index") with open("./metadata.pkl", "rb") as f: docs = pickle.load(f) st_model = SentenceTransformer('all-MiniLM-L6-v2') def make_system_prompt(suffix: str) -> str: return ( "You are EstateGuru, a real estate expert created by Abhishek Pathak from SwavishTek. " "Your role is to help customers buy properties using the available data. " "Only use the provided data—do not make up any information. " "The default currency is AED. If a query uses a different currency, convert the amount to AED " "(for example, $10k becomes 36726.50 AED and $1 becomes 3.67 AED). " "If a customer is interested in a property, wants to buy, or needs to contact an agent or customer care, " "instruct them to call +91 8766268285." f"\n{suffix}" ) general_query_prompt = make_system_prompt( "You are EstateGuru, a helpful real estate assistant. Answer the user's query accurately using the available data. " "Do not invent any details or go beyond the real estate domain. " "If the user shows interest in a property or contacting an agent, ask them to call +91 8766268285." ) # ------------------------ Tool Definitions ------------------------ @tool def extract_filters(query: str) -> dict: """For extracting filters""" # Use a non-streaming ChatQwen for tool use. llm_local = ChatQwen(temperature=0.3, streaming=False) system = ( "You are an expert in extracting filters from property-related queries. Your task is to extract and return only the keys explicitly mentioned in the query as a valid JSON object (starting with '{' and ending with '}'). Include only those keys that are directly present in the query.\n\n" "The possible keys are:\n" " - 'projectName': The name of the project.\n" " - 'developerName': The developer's name.\n" " - 'relationshipManager': The relationship manager.\n" " - 'propertyAddress': The property address.\n" " - 'surroundingArea': The area or nearby landmarks.\n" " - 'propertyType': The type or configuration of the property.\n" " - 'amenities': Any amenities mentioned.\n" " - 'coveredParking': Parking availability.\n" " - 'petRules': Pet policies.\n" " - 'security': Security details.\n" " - 'occupancyRate': Occupancy information.\n" " - 'constructionImpact': Construction or its impact.\n" " - 'propertySize': Size of the property.\n" " - 'propertyView': View details.\n" " - 'propertyCondition': Condition of the property.\n" " - 'serviceCharges': Service or maintenance charges.\n" " - 'ownershipType': Ownership type.\n" " - 'totalCosts': A cost threshold or cost amount.\n" " - 'paymentPlans': Payment or financing plans.\n" " - 'expectedRentalYield': Expected rental yield.\n" " - 'rentalHistory': Rental history.\n" " - 'shortTermRentals': Short-term rental information.\n" " - 'resalePotential': Resale potential.\n" " - 'uniqueId': A unique identifier.\n\n" "Important instructions regarding cost thresholds:\n" " - If the query contains phrases like 'under 10k', 'below 2m', or 'less than 5k', interpret these as cost thresholds.\n" " - Convert any shorthand cost values to pure numbers (for example, '10k' becomes 10000, '2m' becomes 2000000) and assign them to the key 'totalCosts'.\n" " - Do not use 'propertySize' for cost thresholds.\n\n" " - Default currency is AED, if user query have different currency symbol then convert to equivalent AED amount (eg. $10k becomes 36726.50, $1 becomes 3.67).\n\n" "Example:\n" " For the query: \"properties near dubai mall under 43k\"\n" " The expected output should be:\n" " { \"surroundingArea\": \"dubai mall\", \"totalCosts\": 43000 }\n\n" "Return ONLY a valid JSON object with the extracted keys and their corresponding values, with no additional text." ) human_str = f"Here is the query:\n{query}" filter_prompt = [ {"role": "system", "content": system}, {"role": "user", "content": human_str}, ] response = llm_local.invoke(messages=filter_prompt) response_text = response.content if isinstance(response, AIMessage) else str(response) try: model_filters = extract_json_from_response(response_text) except Exception as e: print(f"JSON parsing error: {e}") model_filters = {} rule_filters = rule_based_extract(query) print("Rule-based extraction:", rule_filters) final_filters = {**model_filters, **rule_filters} print("Final extraction:", final_filters) return {"filters": final_filters} @tool def determine_route(query: str) -> dict: """For determining route using enhanced prompt and fallback logic.""" # Define a set of keywords that are strong indicators of a real estate query. real_estate_keywords = estateKeywords # Check if the query includes any of the positive signals. pattern = re.compile("|".join(re.escape(keyword) for keyword in real_estate_keywords), re.IGNORECASE) positive_signal = bool(pattern.search(query)) # Proceed with LLM classification regardless, but use the positive signal in fallback. llm_local = ChatQwen(temperature=0.3, streaming=False) transform_suggest_to_list = query.lower().replace("suggest ", "list ", -1) system = """ Classify the user query as: - **"search"**: if it requests property listings with specific filters (e.g., location, price, property type like "2bhk", service charges, pet policies, etc.). - **"suggest"**: if it asks for property suggestions without filters. - **"detail"**: if it is asking for more information about a previously provided property (for example, "tell me more about property 5" or "I want more information regarding 4BHK"). - **"general"**: for all other real estate-related questions. - **"out_of_domain"**: if the query is not related to real estate (for example, tourist attractions, restaurants, etc.). Keep in mind that queries mentioning terms like "service charge", "allow pets", "pet rules", etc., are considered real estate queries. Return only the keyword: search, suggest, detail, general, or out_of_domain. """ human_str = f"Here is the query:\n{transform_suggest_to_list}" router_prompt = [ {"role": "system", "content": system}, {"role": "user", "content": human_str}, ] response = llm_local.invoke(messages=router_prompt) response_text = response.content if isinstance(response, AIMessage) else str(response) route_value = str(response_text).strip().lower() # Fallback: if the query seems like a detailed request, override. detail_phrases = [ "more information", "tell me more", "more details", "give me more details", "i need more details", "can you provide more details", "additional details", "further information", "expand on that", "explain further", "elaborate more", "more specifics", "i want to know more", "could you elaborate", "need more info", "provide more details", "detail it further", "in-depth information", "break it down further", "further explanation" ] if any(phrase in query.lower() for phrase in detail_phrases): route_value = "detail" if route_value not in {"search", "suggest", "detail", "general", "out_of_domain"}: route_value = "general" if route_value == "out_of_domain" and positive_signal: route_value = "general" if route_value == "out_of_domain": route_value = "general" if positive_signal else "out_of_domain" return {"route": route_value} # ------------------------ Workflow Setup ------------------------ workflow = StateGraph(state_schema=dict) def route_query(state: dict) -> dict: new_state = state.copy() try: new_state["route"] = determine_route.invoke(new_state.get("query", "")).get("route", "general") print(new_state["route"]) except Exception as e: print(f"Routing error: {e}") new_state["route"] = "general" return new_state def hybrid_extract(state: dict) -> dict: new_state = state.copy() new_state["filters"] = extract_filters.invoke(new_state.get("query", "")).get("filters", {}) return new_state def search_faiss(state: dict) -> dict: new_state = state.copy() query_embedding = st_model.encode([state["query"]]) _, indices = index.search(query_embedding.astype(np.float32), 5) new_state["faiss_results"] = [docs[idx] for idx in indices[0] if idx < len(docs)] return new_state def apply_filters(state: dict) -> dict: new_state = state.copy() new_state["final_results"] = apply_filters_partial(state["faiss_results"], state.get("filters", {})) return new_state def suggest_properties(state: dict) -> dict: new_state = state.copy() new_state["suggestions"] = random.sample(docs, 5) return new_state def handle_out_of_domain(state: dict) -> dict: new_state = state.copy() new_state["response"] = "I only handle real estate inquiries. Please ask a question related to properties." return new_state def generate_response(state: dict) -> dict: new_state = state.copy() messages = [] # Add the general query prompt. messages.append({"role": "system", "content": general_query_prompt}) # If this is a detail query, add a system message that forces a detailed answer. if new_state.get("route", "general") == "detail": messages.append({ "role": "system", "content": ( "This is a detail query. Please provide detailed information about the property below. " "Do not generate a new list of properties; only use the provided property details to answer the query. " "Focus on answering the specific question (for example, whether pets are allowed)." ) }) # If property details are available, add them without clearing context. if new_state.get("current_properties"): property_context = format_property_data(new_state["current_properties"]) messages.append({"role": "system", "content": "Available Property:\n" + property_context}) # Do NOT clear current_properties here. messages.append({"role": "system", "content": "When responding, use only the provided property details to answer the user's specific question about the property."}) # Add the conversation history. for msg in state.get("messages", []): if msg["role"] == "user": messages.append({"role": "user", "content": msg["content"]}) else: messages.append({"role": "assistant", "content": msg["content"]}) # Invoke the LLM with the constructed messages. connection_id = state.get("connection_id") loop = state.get("loop") if connection_id and loop: print("Yes") callback_manager = [WebSocketStreamingCallbackHandler(connection_id, loop)] _ = llm.invoke( messages, config={"callbacks": callback_manager} ) new_state["response"] = "" else: callback_manager = [StreamingStdOutCallbackHandler()] response = llm.invoke( messages, config={"callbacks": callback_manager} ) new_state["response"] = response.content if isinstance(response, AIMessage) else str(response) return new_state def format_final_response(state: dict) -> dict: new_state = state.copy() # Only override the current_properties if this is NOT a detail query. if not state.get("route", "general") == "detail": if state.get("route") in ["search", "suggest"]: if "final_results" in state: new_state["current_properties"] = state["final_results"] elif "suggestions" in state: new_state["current_properties"] = state["suggestions"] # Then format the response based on the (possibly filtered) current_properties. if new_state.get("current_properties"): formatted = [] for idx, prop in enumerate(new_state["current_properties"], 1): cost = prop.get("totalCosts", "N/A") cost_str = f"{cost:,}" if isinstance(cost, (int, float)) else cost formatted.append( f"{idx}. Type: {prop['propertyType']}, Cost: AED {cost_str}, " f"Size: {prop.get('propertySize', 'N/A')}, Amenities: {', '.join(map(str, prop.get('amenities', []))) if prop.get('amenities') else 'N/A'}, " f"Rental Yield: {prop.get('expectedRentalYield', 'N/A')}, " f"Ownership: {prop.get('ownershipType', 'N/A')}\n" ) aggregated_response = "Here are the property details:\n" + "\n".join(formatted) connection_id = state.get("connection_id") loop = state.get("loop") if connection_id and loop: import time tokens = aggregated_response.split(" ") for token in tokens: asyncio.run_coroutine_threadsafe( manager_socket.send_message(connection_id, token + " "), loop ) time.sleep(0.05) new_state["response"] = "" else: new_state["response"] = aggregated_response elif "response" in new_state: new_state["response"] = str(new_state["response"]) return new_state nodes = [ ("route_query", route_query), ("hybrid_extract", hybrid_extract), ("faiss_search", search_faiss), ("apply_filters", apply_filters), ("suggest_properties", suggest_properties), ("handle_out_of_domain", handle_out_of_domain), ("generate_response", generate_response), ("format_response", format_final_response) ] for name, node in nodes: workflow.add_node(name, node) workflow.add_edge(START, "route_query") workflow.add_conditional_edges( "route_query", lambda state: state.get("route", "general"), { "search": "hybrid_extract", "suggest": "suggest_properties", "detail": "generate_response", "general": "generate_response", "out_of_domain": "handle_out_of_domain" } ) workflow.add_edge("hybrid_extract", "faiss_search") workflow.add_edge("faiss_search", "apply_filters") workflow.add_edge("apply_filters", "format_response") workflow.add_edge("suggest_properties", "format_response") workflow.add_edge("generate_response", "format_response") workflow.add_edge("handle_out_of_domain", "format_response") workflow.add_edge("format_response", END) workflow_app = workflow.compile() # ------------------------ Conversation Manager ------------------------ class ConversationManager: def __init__(self): self.conversation_history = [] self.current_properties = [] def _add_message(self, role: str, content: str): self.conversation_history.append({ "role": role, "content": content, "timestamp": datetime.now().isoformat() }) def process_query(self, query: str) -> str: # Reset context on greetings to avoid using off-domain history if query.strip().lower() in {"hi", "hello", "hey"}: self.conversation_history = [] self.current_properties = [] greeting_response = "Hello! How can I assist you today with your real estate inquiries?" self._add_message("assistant", greeting_response) return greeting_response try: self._add_message("user", query) initial_state = { "messages": self.conversation_history.copy(), "query": query, "route": "general", "filters": {}, "current_properties": self.current_properties } for event in workflow_app.stream(initial_state, stream_mode="values"): final_state = event if 'final_results' in final_state: self.current_properties = final_state['final_results'] elif 'suggestions' in final_state: self.current_properties = final_state['suggestions'] if final_state.get("route") == "general": response_text = final_state.get("response", "") self._add_message("assistant", response_text) return response_text else: response = final_state.get("response", "I couldn't process that request.") self._add_message("assistant", response) return response except Exception as e: print(f"Processing error: {e}") return "Sorry, I encountered an error processing your request." conversation_managers = {} # ------------------------ FastAPI Backend with WebSockets ------------------------ app = FastAPI() class ConnectionManager: def __init__(self): self.active_connections = {} async def connect(self, websocket: WebSocket): await websocket.accept() connection_id = str(uuid.uuid4()) self.active_connections[connection_id] = websocket print(f"New connection: {connection_id}") return connection_id def disconnect(self, connection_id: str): if connection_id in self.active_connections: del self.active_connections[connection_id] print(f"Disconnected: {connection_id}") async def send_message(self, connection_id: str, message: str): websocket = self.active_connections.get(connection_id) if websocket: await websocket.send_text(message) manager_socket = ConnectionManager() def stream_query(query: str, connection_id: str, loop): conv_manager = conversation_managers.get(connection_id) if conv_manager is None: print(f"No conversation manager found for connection {connection_id}") return # Check for greetings and handle them immediately if query.strip().lower() in {"hi", "hello", "hey"}: conv_manager.conversation_history = [] conv_manager.current_properties = [] greeting_response = "Hello! How can I assist you today with your real estate inquiries?" conv_manager._add_message("assistant", greeting_response) asyncio.run_coroutine_threadsafe( manager_socket.send_message(connection_id, greeting_response), loop ) return conv_manager._add_message("user", query) initial_state = { "messages": conv_manager.conversation_history.copy(), "query": query, "route": "general", "filters": {}, "current_properties": conv_manager.current_properties, "connection_id": connection_id, "loop": loop } try: workflow_app.invoke(initial_state) except Exception as e: error_msg = f"Error processing query: {str(e)}" asyncio.run_coroutine_threadsafe( manager_socket.send_message(connection_id, error_msg), loop ) @app.websocket("/ws") async def websocket_endpoint(websocket: WebSocket): connection_id = await manager_socket.connect(websocket) conversation_managers[connection_id] = ConversationManager() try: while True: query = await websocket.receive_text() loop = asyncio.get_event_loop() # loop = asyncio.get_running_loop() threading.Thread( target=stream_query, args=(query, connection_id, loop), daemon=True ).start() except WebSocketDisconnect: conv_manager = conversation_managers.get(connection_id) if conv_manager: filename = f"conversations/conversation_{connection_id}_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json" with open(filename, "w") as f: json.dump(conv_manager.conversation_history, f, indent=4) del conversation_managers[connection_id] manager_socket.disconnect(connection_id) @app.post("/query") async def post_query(query: str): conv_manager = ConversationManager() response = conv_manager.process_query(query) return {"response": response}