import gradio as gr from langchain_mistralai.chat_models import ChatMistralAI from langchain.prompts import ChatPromptTemplate import os from pathlib import Path import json import faiss import numpy as np from langchain.schema import Document import pickle import re import requests from functools import lru_cache import torch from sentence_transformers import SentenceTransformer from sentence_transformers.cross_encoder import CrossEncoder import threading from queue import Queue import concurrent.futures from typing import Generator, Tuple import time class OptimizedRAGLoader: def __init__(self, docs_folder: str = "./docs", splits_folder: str = "./splits", index_folder: str = "./index"): self.docs_folder = Path(docs_folder) self.splits_folder = Path(splits_folder) self.index_folder = Path(index_folder) # Create folders if they don't exist for folder in [self.splits_folder, self.index_folder]: folder.mkdir(parents=True, exist_ok=True) # File paths self.splits_path = self.splits_folder / "splits.json" self.index_path = self.index_folder / "faiss.index" self.documents_path = self.index_folder / "documents.pkl" # Initialize components self.index = None self.indexed_documents = None # Initialize encoder model self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu") self.encoder = SentenceTransformer("intfloat/multilingual-e5-large") self.encoder.to(self.device) self.reranker = model = CrossEncoder("cross-encoder/mmarco-mMiniLMv2-L12-H384-v1",trust_remote_code=True) # Initialize thread pool self.executor = concurrent.futures.ThreadPoolExecutor(max_workers=4) # Initialize response cache self.response_cache = {} @lru_cache(maxsize=1000) def encode(self, text: str): """Cached encoding function""" with torch.no_grad(): embeddings = self.encoder.encode( text, convert_to_numpy=True, normalize_embeddings=True ) return embeddings def batch_encode(self, texts: list): """Batch encoding for multiple texts""" with torch.no_grad(): embeddings = self.encoder.encode( texts, batch_size=32, convert_to_numpy=True, normalize_embeddings=True, show_progress_bar=False ) return embeddings def load_and_split_texts(self): if self._splits_exist(): return self._load_existing_splits() documents = [] futures = [] for file_path in self.docs_folder.glob("*.txt"): future = self.executor.submit(self._process_file, file_path) futures.append(future) for future in concurrent.futures.as_completed(futures): documents.extend(future.result()) self._save_splits(documents) return documents def _process_file(self, file_path): with open(file_path, 'r', encoding='utf-8') as file: text = file.read() chunks = [s.strip() for s in re.split(r'(?<=[.!?])\s+', text) if s.strip()] return [ Document( page_content=chunk, metadata={ 'source': file_path.name, 'chunk_id': i, 'total_chunks': len(chunks) } ) for i, chunk in enumerate(chunks) ] def load_index(self) -> bool: """ Charge l'index FAISS et les documents associés s'ils existent Returns: bool: True si l'index a été chargé, False sinon """ if not self._index_exists(): print("Aucun index trouvé.") return False print("Chargement de l'index existant...") try: # Charger l'index FAISS self.index = faiss.read_index(str(self.index_path)) # Charger les documents associés with open(self.documents_path, 'rb') as f: self.indexed_documents = pickle.load(f) print(f"Index chargé avec {self.index.ntotal} vecteurs") return True except Exception as e: print(f"Erreur lors du chargement de l'index: {e}") return False def create_index(self, documents=None): if documents is None: documents = self.load_and_split_texts() if not documents: return False texts = [doc.page_content for doc in documents] embeddings = self.batch_encode(texts) dimension = embeddings.shape[1] self.index = faiss.IndexFlatL2(dimension) if torch.cuda.is_available(): # Use GPU for FAISS if available res = faiss.StandardGpuResources() self.index = faiss.index_cpu_to_gpu(res, 0, self.index) self.index.add(np.array(embeddings).astype('float32')) self.indexed_documents = documents # Save index and documents cpu_index = faiss.index_gpu_to_cpu(self.index) if torch.cuda.is_available() else self.index faiss.write_index(cpu_index, str(self.index_path)) with open(self.documents_path, 'wb') as f: pickle.dump(documents, f) return True def _index_exists(self) -> bool: """Vérifie si l'index et les documents associés existent""" return self.index_path.exists() and self.documents_path.exists() def get_retriever(self, k: int = 10): if self.index is None: if not self.load_index(): if not self.create_index(): raise ValueError("Unable to load or create index") def retriever_function(query: str) -> list: # Check cache first cache_key = f"{query}_{k}" if cache_key in self.response_cache: return self.response_cache[cache_key] query_embedding = self.encode(query) distances, indices = self.index.search( np.array([query_embedding]).astype('float32'), k ) results = [ self.indexed_documents[idx] for idx in indices[0] if idx != -1 ] # Cache the results self.response_cache[cache_key] = results return results return retriever_function # Initialize components mistral_api_key = os.getenv("mistral_api_key") llm = ChatMistralAI( model="mistral-large-latest", mistral_api_key=mistral_api_key, temperature=0.01, streaming=True, ) rag_loader = OptimizedRAGLoader() retriever = rag_loader.get_retriever(k=30) # Reduced k for faster retrieval # Cache for processed questions question_cache = {} prompt_template = ChatPromptTemplate.from_messages([ ("system", """أنت مساعد مفيد يجيب على الأسئلة باللغة العربية باستخدام المعلومات المقدمة. استخدم المعلومات التالية للإجابة على السؤال: {context} إذا لم تكن المعلومات كافية للإجابة على السؤال بشكل كامل، قل لا أعرف. أجب بشكل موجز ودقيق. /n أذكر رقم المادة أو الفصل حسب الحالة. أذكر اسم ورقم القانون ان كان متوفرا وان لم يكن لا تذكر شيئا. """), ("human", "{question}") ]) import gradio as gr from typing import Iterator # Ajouter du CSS pour personnaliser l'apparence css = """ /* Reset RTL global */ *, *::before, *::after { direction: rtl !important; text-align: right !important; } body { font-family: 'Amiri', sans-serif; /* Utilisation de la police Arabe andalouse */ background-color: black; /* Fond blanc */ color: black !important; /* Texte noir */ direction: rtl !important; /* Texte en arabe aligné à droite */ } .gradio-container { direction: rtl !important; /* Alignement RTL pour toute l'interface */ } /* Éléments de formulaire */ input[type="text"], .gradio-textbox input, textarea { border-radius: 20px; padding: 10px 15px; border: 2px solid #000; font-size: 16px; width: 80%; margin: 0 auto; text-align: right !important; } /* Surcharge des styles de placeholder */ input::placeholder, textarea::placeholder { text-align: right !important; direction: rtl !important; } /* Boutons */ .gradio-button { border-radius: 20px; font-size: 16px; background-color: #007BFF; color: white; padding: 10px 20px; margin: 10px auto; border: none; width: 80%; display: block; } .gradio-button:hover { background-color: #0056b3; } .gradio-chatbot .message { border-radius: 20px; padding: 10px; margin: 10px 0; background-color: #f1f1f1; border: 1px solid #ddd; width: 80%; text-align: right !important; direction: rtl !important; } /* Messages utilisateur alignés à gauche */ .gradio-chatbot .user-message { margin-right: auto; background-color: #e3f2fd; text-align: right !important; direction: rtl !important; } /* Messages assistant alignés à droite */ .gradio-chatbot .assistant-message { margin-right: auto; background-color: #f1f1f1; text-align: right } /* Corrections RTL pour les éléments spécifiques */ .gradio-textbox textarea { text-align: right !important; } .gradio-dropdown div { text-align: right !important; } """ def process_question(question: str) -> Iterator[str]: if question in question_cache: response, docs = question_cache[question] yield response + "\nSources:\n" + "\n".join([doc.page_content for doc in docs]) return relevant_docs = retriever(question) # context = "\n".join([doc.page_content for doc in relevant_docs]) context = [doc.page_content for doc in relevant_docs] text_pairs = [[question, text] for text in context] scores = rag_loader.reranker.predict(text_pairs) scored_docs = list(zip(scores, context, relevant_docs)) # scored_docs.sort(reverse=True) scored_docs.sort(key=lambda x: x[0], reverse=True) reranked_docs = [d[2].page_content for d in scored_docs][:6] prompt = prompt_template.format_messages( context=reranked_docs, question=question ) full_response = "" try: for chunk in llm.stream(prompt): if isinstance(chunk, str): current_chunk = chunk else: current_chunk = chunk.content full_response += current_chunk # sources = "\n".join(set([doc.metadata.get("source") for doc in relevant_docs])) # sources = [os.path.splitext(source[1])[0] for source in sources] # yield full_response + "\n\n\nالمصادر المحتملة :\n" + "".join(sources) sources = [doc.metadata.get("source") for doc in relevant_docs] sources = list(set([os.path.splitext(source)[0] for source in sources])) sources = [d[2].metadata['source'] for d in scored_docs][:6] sources = list(set([os.path.splitext(source)[0] for source in sources])) yield full_response + "\n\n\nالمصادر المحتملة :\n" + "\n".join(sources) # yield full_response + "\n\n\nالمصادر المحتملة:\n" + "\n".join([doc.metadata.get("source") for doc in relevant_docs]) question_cache[question] = (full_response, relevant_docs) except Exception as e: yield f"Erreur lors du traitement : {str(e)}" # def process_question(question: str) -> Iterator[str]: # """ # Process the question and return a response generator for streaming. # """ # if question in question_cache: # yield question_cache[question][0] # return # relevant_docs = retriever(question) # context = "\n".join([doc.page_content for doc in relevant_docs]) # prompt = prompt_template.format_messages( # context=context, # question=question # ) # full_response = "" # try: # for chunk in llm.stream(prompt): # if isinstance(chunk, str): # current_chunk = chunk # else: # current_chunk = chunk.content # full_response += current_chunk # yield full_response # Send the updated response in streaming # question_cache[question] = (full_response, context) # except Exception as e: # yield f"Erreur lors du traitement : {str(e)}" # def process_question(question: str) -> Iterator[str]: # """ # Process the question and return a response generator for streaming, including sources. # """ # if question in question_cache: # yield question_cache[question][0] # return # # Récupérer les documents pertinents # relevant_docs = retriever(question) # context = "\n".join([doc.page_content for doc in relevant_docs]) # sources = [doc.metadata.get("source", "Source inconnue") for doc in relevant_docs] # sources = os.path.splitext(sources[0])[0] if sources else "غير معروف" # # Générer le prompt # prompt = prompt_template.format_messages( # context=context, # question=question # ) # full_response = "" # try: # # Streaming de la réponse # for chunk in llm.stream(prompt): # if isinstance(chunk, str): # current_chunk = chunk # else: # current_chunk = chunk.content # full_response += current_chunk # yield full_response # Envoyer la réponse mise à jour en streaming # # Ajouter les sources à la réponse finale # if sources: # sources_str = "\nSources :\n" + "\n".join(f"- {source}" for source in sources) # full_response += sources_str # yield sources_str # Envoyer les sources # # Mettre en cache la réponse complète # question_cache[question] = (full_response, context) # except Exception as e: # yield f"Erreur lors du traitement : {str(e)}" # def process_question(question: str) -> tuple[str, list[str]]: # # Check cache first # if question in question_cache: # return question_cache[question] # # Get relevant documents using the retriever # relevant_docs = retriever(question) # # Extract the content and sources # context = "\n".join(doc.page_content for doc in relevant_docs) # sources = [doc.metadata["source"] for doc in relevant_docs] # sources = os.path.splitext(sources[0])[0] if sources else "غير معروف" # # Generate the prompt with the context # prompt = prompt_template.format( # context=context, # question=question # ) # full_response = "" # try: # for chunk in llm.stream(prompt): # if isinstance(chunk, str): # current_chunk = chunk # else: # current_chunk = chunk.content # full_response += current_chunk # yield full_response, sources # Send the updated response in streaming # question_cache[question,sources] = (full_response, context) # except Exception as e: # yield f"Erreur lors du traitement : {str(e)}" def gradio_stream(question: str, chat_history: list) -> Iterator[list]: """ Format the output for Gradio Chatbot component with streaming. """ full_response = "" try: for partial_response in process_question(question): full_response = partial_response # Append the latest assistant response to chat history updated_chat = chat_history + [[question, partial_response]] yield updated_chat except Exception as e: # Handle errors during streaming updated_chat = chat_history + [[question, f"Erreur : {str(e)}"]] yield updated_chat # def gradio_stream(question: str, chat_history: list) -> Iterator[list]: # """ # Format the output for Gradio Chatbot component with streaming, including sources. # """ # full_response = "" # sources_str = "" # try: # for partial_response in process_question(question): # if "Sources :" in partial_response: # # Les sources sont ajoutées à la réponse finale # sources_str = partial_response # updated_chat = chat_history + [[question, full_response + "\n" + sources_str]] # else: # # Construire progressivement la réponse # full_response = partial_response # updated_chat = chat_history + [[question, full_response]] # yield updated_chat # except Exception as e: # # Gestion des erreurs lors du streaming # updated_chat = chat_history + [[question, f"Erreur : {str(e)}"]] # yield updated_chat # Gradio interface with gr.Blocks(css=css) as demo: gr.Markdown("

هذا تطبيق للاجابة على الأسئلة المتعلقة يالقوانين المغربية

") # Organisation en 3 lignes with gr.Row(): # Première ligne: Question message = gr.Textbox(label="أدخل سؤالك", placeholder="اكتب سؤالك هنا", elem_id="question_input") with gr.Row(): # Deuxième ligne: Bouton de recherche send = gr.Button("بحث", elem_id="search_button") with gr.Row(): # Troisième ligne: Affichage de la réponse chatbot = gr.Chatbot(label="") # Fonction de mise à jour pour l'utilisateur def user_input(user_message, chat_history): return "", chat_history + [[user_message, None]] send.click(user_input, [message, chatbot], [message, chatbot], queue=False) send.click(gradio_stream, [message, chatbot], chatbot) demo.launch(share=True) # # def process_question(question: str): # # """ # # Process the question and yield the answer progressively. # # """ # # # Check cache first # # if question in question_cache: # # yield question_cache[question] # Retourne directement depuis le cache si disponible # # relevant_docs = retriever(question) # # context = "\n".join([doc.page_content for doc in relevant_docs]) # # prompt = prompt_template.format_messages( # # context=context, # # question=question # # ) # # response = "" # Initialise la réponse # # # Ici, nous supposons que 'llm.stream' est un générateur qui renvoie des chunks # # for chunk in llm.stream(prompt): # suppose que llm.stream renvoie des chunks de réponse # # if isinstance(chunk, str): # # response += chunk # Accumulez la réponse si c'est déjà une chaîne # # else: # # response += chunk.content # Sinon, prenez le contenu du chunk (si chunk est un type d'objet spécifique) # # yield response, context # Renvoie la réponse mise à jour et le contexte # # # Mettez le résultat en cache à la fin # # # question_cache[question] = (response, context) # def process_question(question: str) -> Generator[Tuple[str, str], None, None]: # """ # Process the question and yield the answer progressively. # """ # # Check cache first # if question in question_cache: # yield question_cache[question] # relevant_docs = retriever(question) # context = "\n".join([doc.page_content for doc in relevant_docs]) # prompt = prompt_template.format_messages( # context=context, # question=question # ) # current_response = "" # for chunk in llm.stream(prompt): # if isinstance(chunk, str): # current_response += chunk # else: # current_response += chunk.content # yield current_response, context # # Mettez le résultat en cache à la fin # question_cache[question] = (response, context) # # CSS personnalisé avec l'importation de Google Fonts # custom_css = """ # /* Import Google Fonts - Noto Sans Arabic */ # @import url('https://fonts.googleapis.com/css2?family=Noto+Sans+Arabic:wght@300;400;500;600;700&display=swap'); # /* Styles généraux */ # :root { # --primary-color: #4299e1; # --secondary-color: #666666; # --accent-color: #4299E1; # --background-color: #ffffff; # --border-radius: 8px; # --font-family-arabic: 'Noto Sans Arabic', Arial, sans-serif; # } # /* Style de base */ # body { # font-family: var(--font-family-arabic); # background-color: var(--background-color); # color: var(--primary-color); # } # /* Styles pour le texte RTL */ # .rtl-text { # text-align: right !important; # direction: rtl !important; # font-family: var(--font-family-arabic) !important; # } # .rtl-text textarea { # text-align: right !important; # direction: rtl !important; # padding: 1rem !important; # border-radius: var(--border-radius) !important; # border: 1px solid #E2E8F0 !important; # background-color: #ffffff !important; # color: var(--primary-color) !important; # font-size: 1.1rem !important; # line-height: 1.6 !important; # font-family: var(--font-family-arabic) !important; # } # /* Style du titre */ # .app-title { # font-family: var(--font-family-arabic) !important; # font-size: 2rem !important; # font-weight: 700 !important; # color: white !important; /* Texte en blanc */ # background-color: #3e2b1f !important; /* Fond marron foncé */ # margin-bottom: 1rem !important; # margin-top: 1rem !important; # text-align: center !important; # } # /* Styles des étiquettes */ # .rtl-text label { # font-family: var(--font-family-arabic) !important; # font-size: 1.2rem !important; # font-weight: 600 !important; # color: #000000 !important; /* Couleur noire pour les étiquettes */ # margin-bottom: 0.5rem !important; # } # /* Centrer le bouton */ # button.primary-button { # font-family: var(--font-family-arabic) !important; # background-color: var(--accent-color) !important; # color: white !important; # padding: 0.75rem 1.5rem !important; # border-radius: var(--border-radius) !important; # font-weight: 600 !important; # font-size: 1.1rem !important; # transition: all 0.3s ease !important; # width: 200px !important; /* Réduit la largeur du bouton */ # margin: 0 auto !important; /* Centrage horizontal */ # display: block !important; /* Nécessaire pour que le margin auto fonctionne */ # } # button.primary-button:hover { # background-color: #3182CE !important; # transform: translateY(-1px) !important; # } # /* Styles des boîtes de texte */ # .textbox-container { # background-color: #b45f06 !important; # padding: 1.5rem !important; # border-radius: var(--border-radius) !important; # box-shadow: 0 2px 4px rgba(0, 0, 0, 0.1) !important; # margin-bottom: 1rem !important; # } # /* Animation de chargement */ # .loading { # animation: pulse 2s infinite; # } # @keyframes pulse { # 0% { opacity: 1; } # 50% { opacity: 0.5; } # 100% { opacity: 1; } # } # /* Style du statut */ # .status-text { # font-family: var(--font-family-arabic) !important; # text-align: center !important; # color: var(--secondary-color) !important; # font-size: 1rem !important; # margin-top: 1rem !important; # } # """ # # Interface Gradio avec streaming # with gr.Blocks(css=custom_css) as iface: # with gr.Column(elem_classes="container"): # gr.Markdown( # "# نظام الأسئلة والأجوبة الذكي", # elem_classes="app-title rtl-text" # ) # with gr.Column(elem_classes="textbox-container"): # input_text = gr.Textbox( # label="السؤال", # placeholder="اكتب سؤالك هنا...", # lines=1, # elem_classes="rtl-text" # ) # with gr.Row(): # with gr.Column(): # answer_box = gr.Textbox( # label="الإجابة", # lines=5, # elem_classes="rtl-text textbox-container" # ) # submit_btn = gr.Button( # "إرسال السؤال", # elem_classes="primary-button", # variant="primary" # ) # # def stream_response(question): # # response_stream = process_question(question) # # for response, _ in response_stream: # # gr.update(value=response) # # yield response # def stream_response(question): # for chunk_response, _ in process_question(question): # yield chunk_response # time.sleep(0.05) # submit_btn.click( # fn=stream_response, # inputs=input_text, # outputs=answer_box, # api_name="predict", # queue=False # ) # if __name__ == "__main__": # iface.launch( # share=True, # server_name="0.0.0.0", # server_port=7860, # max_threads=3, # show_error=True # ) # # # Interface Gradio avec la correction # # with gr.Blocks(css=custom_css) as iface: # # with gr.Column(elem_classes="container"): # # gr.Markdown( # # "# نظام الأسئلة والأجوبة الذكي", # # elem_classes="app-title rtl-text" # # ) # # with gr.Column(elem_classes="textbox-container"): # # input_text = gr.Textbox( # # label="السؤال", # # placeholder="اكتب سؤالك هنا...", # # lines=1, # # elem_classes="rtl-text" # # ) # # with gr.Row(): # # with gr.Column(): # # answer_box = gr.Textbox( # # label="الإجابة", # # lines=5, # # elem_classes="rtl-text textbox-container" # # ) # # # with gr.Column(scale=1): # # # context_box = gr.Textbox( # # # label="السياق المستخدم", # # # lines=4, # # # elem_classes="rtl-text textbox-container" # # # ) # # submit_btn = gr.Button( # # "إرسال السؤال", # # elem_classes="primary-button", # # variant="primary" # # ) # # def on_submit(question): # # for response, context in process_question(question): # # yield response # Yield À CHAQUE itération # # submit_btn.click( # # fn=on_submit, # # inputs=input_text, # # outputs=answer_box, # # api_name="predict", # # queue=False, # # ) # # if __name__ == "__main__": # # iface.launch( # # share=True, # # server_name="0.0.0.0", # # server_port=7860, # # max_threads=3, # # show_error=True, # # ) # # def process_question(question: str): # # """ # # Process the question and yield the answer progressively. # # """ # # # Check cache first # # if question in question_cache: # # yield question_cache[question] # Retourne directement depuis le cache si disponible # # relevant_docs = retriever(question) # # context = "\n".join([doc.page_content for doc in relevant_docs]) # # prompt = prompt_template.format_messages( # # context=context, # # question=question # # ) # # response = "" # Initialise la réponse # # # Ici, nous supposons que 'llm.stream' est un générateur qui renvoie des chunks # # for chunk in llm.stream(prompt): # suppose que llm.stream renvoie des chunks de réponse # # if isinstance(chunk, str): # # response += chunk # Accumulez la réponse si c'est déjà une chaîne # # else: # # response += chunk.content # Sinon, prenez le contenu du chunk (si chunk est un type d'objet spécifique) # # yield response, context # Renvoie la réponse mise à jour et le contexte # # # Mettez le résultat en cache à la fin # # question_cache[question] = (response, context) # # # Custom CSS for right-aligned text in textboxes # # custom_css = """ # # .rtl-text { # # text-align: right !important; # # direction: rtl !important; # # } # # .rtl-text textarea { # # text-align: right !important; # # direction: rtl !important; # # } # # """ # # # Gradio interface with queue # # with gr.Blocks(css=custom_css) as iface: # # with gr.Column(): # # input_text = gr.Textbox( # # label="السؤال", # # placeholder="اكتب سؤالك هنا...", # # lines=2, # # elem_classes="rtl-text" # # ) # # with gr.Row(): # # answer_box = gr.Textbox( # # label="الإجابة", # # lines=4, # # elem_classes="rtl-text" # # ) # # context_box = gr.Textbox( # # label="السياق المستخدم", # # lines=8, # # elem_classes="rtl-text" # # ) # # submit_btn = gr.Button("إرسال") # # submit_btn.click( # # fn=process_question, # # inputs=input_text, # # outputs=[answer_box, context_box], # # api_name="predict", # # queue=True # Utiliser le système de queue pour un traitement asynchrone # # ) # # if __name__ == "__main__": # # iface.launch( # # share=True, # # server_name="0.0.0.0", # # server_port=7860, # # max_threads=3, # Controls concurrency # # show_error=True # # ) # # def process_question(question: str): # # """ # # Process the question and return the answer and context # # """ # # # Check cache first # # if question in question_cache: # # return question_cache[question], "" # Retourne la réponse cachée et un statut vide # # relevant_docs = retriever(question) # # context = "\n".join([doc.page_content for doc in relevant_docs]) # # prompt = prompt_template.format_messages( # # context=context, # # question=question # # ) # # response = "" # # for chunk in llm.stream(prompt): # # if isinstance(chunk, str): # # response += chunk # # else: # # response += chunk.content # # # Mettez le résultat en cache à la fin # # question_cache[question] = (response, context) # # return response, context