src/agent.py

# # This file will contain the core logic for the TurBot agent,
# # including the LangGraph setup, state management, and tool definition. 
# 
# import os
# from typing import List
# from dotenv import load_dotenv
# from huggingface_hub.inference._client import InferenceClient
# from langchain_community.vectorstores import FAISS
# from langchain_huggingface import HuggingFaceEmbeddings
# from langchain_core.messages import HumanMessage, AIMessage
# 
# # Load environment variables
# load_dotenv()
# 
# # Initialize the LLM (using a free model from Hugging Face)
# client = InferenceClient("meta-llama/Meta-Llama-3-70B-Instruct")
# 
# # Load the vector store
# DB_FAISS_PATH = "../vectorstore/db_faiss"
# 
# def search_travel_packages(query: str) -> str:
#     """
#     Search for travel packages based on the user's query.
#     This function searches through the travel package documents to find relevant information.
#     """
#     print(f"BURAAAZ STA SAM JA USAO U SEARCH TRAVEL PACKAGES OVO STVARNO NISAM OCEKIVAO")
#     try:
#         # Load the vector store
#         embeddings = HuggingFaceEmbeddings(
#             model_name='sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2',
#             model_kwargs={'device': 'cpu'}
#         )
#         
#         if os.path.exists(DB_FAISS_PATH):
#             vectorstore = FAISS.load_local(DB_FAISS_PATH, embeddings)
#             
# 
#             print(f"JOOOO USPEJO SAM")
# 
#             # Search for relevant documents
#             docs = vectorstore.similarity_search(query, k=3)
#             
#             # Combine the relevant information
#             context = "\n\n".join([doc.page_content for doc in docs])
#             return f"Found relevant travel information:\n\n{context}"
#         else:
#             return "Travel package database not found. Please ensure the vector store has been created."
#     except Exception as e:
#         return f"Error searching travel packages: {str(e)}"
# 
# def should_search_travel_info(message: str) -> bool:
#     print(f"BURAZENGIJA SRA CU JA U SHOULD SEARCH TRAVEL INFU")
#     """Decide whether to search for travel information based on the message content."""
#     message_lower = message.lower()
#     
#     # Keywords that indicate we should search for travel information
#     travel_keywords = [
#         "destinacija", "let", "hotel", "putovanje", "mediteran", "grčka", "italija", 
#         "turska", "budžet", "porodica", "all inclusive", "preporučite", "ponuda",
#         "destination", "flight", "travel", "mediterranean", "greece", "italy", 
#         "turkey", "budget", "family", "recommend", "offer"
#     ]
#     
#     return any(keyword in message_lower for keyword in travel_keywords)
# 
# def chat_with_turbot(message: str, history: List[List[str]]) -> str:
#     print(f"BURAZ KOJI JE OVO KITA MOJA")
#     """
#     Main function to chat with TurBot.
#     """
#     # Prepare the system message first
#     system_message = """Ti si TurBot, digitalni asistent za turističku agenciju. 
#     Odgovori na srpskom jeziku i budi koristan i prijateljski nastrojen. 
#     Ako ne znaš odgovor na pitanje, preporuči korisniku da pita o putovanjima, letovima, hotelima ili destinacijama."""
#     
#     # Convert messages to the format expected by InferenceClient
#     messages_for_client = []
#     
#     # Add system message first (only once)
#     messages_for_client.append({"role": "system", "content": system_message})
#     
#     # Add conversation history
#     for human, ai in history:
#         if human:
#             messages_for_client.append({"role": "user", "content": human})
#         if ai:
#             messages_for_client.append({"role": "assistant", "content": ai})
#     
#     # Add the current message
#     messages_for_client.append({"role": "user", "content": message})
#     
#     # Check if we should search for travel information
#     context = ""
#     if should_search_travel_info(message):
#         context = search_travel_packages(message)
#         if context:
#             # Add context as additional information
#             messages_for_client.append({"role": "user", "content": f"Kontekst o putovanjima: {context}"})
#     
#     try:
#         # Debug: Print messages being sent to model
#         print(f"\n🔍 DEBUG: Sending {len(messages_for_client)} messages to model:")
#         for i, msg in enumerate(messages_for_client):
#             print(f"  {i+1}. {msg['role']}: {msg['content'][:100]}...")
#         
#         # Generate response
#         response = ""
#         for message_chunk in client.chat_completion(
#             messages_for_client,
#             stream=True,
#             max_tokens=512,
#             temperature=0.7,
#             top_p=0.95,
#         ):
#             if hasattr(message_chunk, 'choices') and message_chunk.choices:
#                 token = message_chunk.choices[0].delta.content
#                 if token:
#                     response += token
#         
#         return response if response else "Izvinjavam se, nisam dobio odgovor od modela. Pokušajte ponovo."
#     except Exception as e:
#         return f"Izvinjavam se, došlo je do greške: {str(e)}" 

import os
from typing import List
from dotenv import load_dotenv
from huggingface_hub.inference._client import InferenceClient
from langchain_community.vectorstores import FAISS
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_core.messages import HumanMessage, AIMessage

# Load environment variables
load_dotenv()

# Initialize the LLM (using a free model from Hugging Face)
client = InferenceClient("meta-llama/Meta-Llama-3-70B-Instruct")

# Load the vector store
DB_FAISS_PATH = "../vectorstore/db_faiss"

# --- KLJUČNA IZMENA: Učitaj embeddings model i FAISS bazu SAMO JEDNOM na početku ---
# Inicijalizuj embeddings model globalno (ili bar van funkcije search_travel_packages)
print("Initializing HuggingFaceEmbeddings...")
try:
    embeddings_model = HuggingFaceEmbeddings(
        model_name='sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2',
        model_kwargs={'device': 'cpu'} # Možeš probati 'cuda' ako imaš GPU i PyTorch sa CUDA
    )
    print("HuggingFaceEmbeddings initialized.")
except Exception as e:
    print(f"Error initializing embeddings: {str(e)}")
    embeddings_model = None # Postavi na None ako dođe do greške

# Učitaj FAISS vektor bazu globalno
vectorstore = None

if embeddings_model:
    print(f"JOOOOO EMBEDINGS POSTOJI BACI MOJ")

if embeddings_model and os.path.exists(DB_FAISS_PATH):
    print(f"Loading FAISS vector store from {DB_FAISS_PATH}...")
    try:
        vectorstore = FAISS.load_local(DB_FAISS_PATH, embeddings_model, allow_dangerous_deserialization=True)
        print("FAISS vector store loaded successfully.")
    except Exception as e:
        print(f"Error loading FAISS vector store: {str(e)}. Make sure 'faiss-cpu' is installed and vectorstore is valid.")
        vectorstore = None
else:
    print("FAISS vector store path does not exist or embeddings model failed to initialize. RAG will not function.")

# --- KRAJ KLJUČNE IZMENE ---


def search_travel_packages(query: str) -> str:
    """
    Search for travel packages based on the user's query.
    This function searches through the travel package documents to find relevant information.
    """
    print(f"BURAAAZ STA SAM JA USAO U SEARCH TRAVEL PACKAGES OVO STVARNO NISAM OCEKIVAO")
    
    if vectorstore is None:
        return "Travel package database is not available. Please ensure the vector store has been created and loaded correctly."
        
    try:
        # Nema više učitavanja embeddings i vectorstore OVDE! Koristimo globalne instance.
        
        print(f"JOOOO USPEJO SAM (now searching)...") # Poruka je sada relevantnija

        # Search for relevant documents
        docs = vectorstore.similarity_search(query, k=3)
        
        # Combine the relevant information
        context = "\n\n".join([doc.page_content for doc in docs])
        return f"Found relevant travel information:\n\n{context}"
    except Exception as e:
        return f"Error searching travel packages: {str(e)}"

def should_search_travel_info(message: str) -> bool:
    print(f"BURAZENGIJA SRA CU JA U SHOULD SEARCH TRAVEL INFU")
    """Decide whether to search for travel information based on the message content."""
    message_lower = message.lower()
    
    # Keywords that indicate we should search for travel information
    travel_keywords = [
        "destinacija", "let", "hotel", "putovanje", "mediteran", "grčka", "italija", 
        "turska", "budžet", "porodica", "all inclusive", "preporučite", "ponuda",
        "destination", "flight", "travel", "mediterranean", "greece", "italy", 
        "turkey", "budget", "family", "recommend", "offer"
    ]
    
    return any(keyword in message_lower for keyword in travel_keywords)

def chat_with_turbot(message: str, history: List[List[str]]) -> str:
    print(f"BURAZ KOJI JE OVO KITA MOJA")
    """
    Main function to chat with TurBot.
    """
    # Prepare the system message first
    system_message = """Ti si TurBot, digitalni asistent za turističku agenciju. 
    Odgovori na srpskom jeziku i budi koristan i prijateljski nastrojen. 
    Ako ne znaš odgovor na pitanje, preporuči korisniku da pita o putovanjima, letovima, hotelima ili destinacijama."""
    
    # Convert messages to the format expected by InferenceClient
    messages_for_client = []
    
    # Add system message first (only once per interaction, model will handle context)
    messages_for_client.append({"role": "system", "content": system_message})
    
    # Add conversation history
    for human, ai in history:
        if human:
            messages_for_client.append({"role": "user", "content": human})
        if ai:
            messages_for_client.append({"role": "assistant", "content": ai})
    
    # Add the current message
    messages_for_client.append({"role": "user", "content": message})
    
    # Check if we should search for travel information
    context = ""
    if should_search_travel_info(message):
        context = search_travel_packages(message)
        if context:
            # Add context as additional information for the model
            # It's better to add it as a user message or a tool output for the model to use
            messages_for_client.append({"role": "user", "content": f"Relevantne informacije iz baze podataka: {context}"})
    
    try:
        # Debug: Print messages being sent to model
        print(f"\n🔍 DEBUG: Sending {len(messages_for_client)} messages to model:")
        for i, msg in enumerate(messages_for_client):
            print(f"   {i+1}. {msg['role']}: {msg['content'][:150]}...") # Povećan prikaz karaktera za debug
        
        # Generate response
        response = ""
        # The `client.chat_completion` now expects `messages` in a list of dicts.
        # This is already correctly prepared as `messages_for_client`.
        for message_chunk in client.chat_completion(
            messages_for_client, # Prosleđujemo celu listu poruka
            stream=True,
            max_tokens=512,
            temperature=0.7,
            top_p=0.95,
        ):
            if hasattr(message_chunk, 'choices') and message_chunk.choices:
                token = message_chunk.choices[0].delta.content
                if token:
                    response += token
        
        return response if response else "Izvinjavam se, nisam dobio odgovor od modela. Pokušajte ponovo."
    except Exception as e:
        return f"Izvinjavam se, došlo je do greške u pozivu LLM-a: {str(e)}"