import gradio as gr from langchain_mistralai.chat_models import ChatMistralAI from langchain.prompts import ChatPromptTemplate from langchain_deepseek import ChatDeepSeek from langchain_google_genai import ChatGoogleGenerativeAI 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, Iterator 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, # # ) # # deepseek_api_key = os.getenv("DEEPSEEK_KEY") # # llm = ChatDeepSeek( # # model="deepseek-chat", # # temperature=0, # # api_key=deepseek_api_key, # # streaming=True, # # ) # gemini_api_key = os.getenv("GEMINI_KEY") # llm = ChatGoogleGenerativeAI( # model="gemini-1.5-pro", # temperature=0, # google_api_key=gemini_api_key, # disable_streaming=True, # ) # rag_loader = OptimizedRAGLoader() # retriever = rag_loader.get_retriever(k=5) # Reduced k for faster retrieval # # Cache for processed questions # question_cache = {} # prompt_template = ChatPromptTemplate.from_messages([ # ("system", """Vous êtes un assistant juridique expert qualifié. Analysez et répondez aux questions juridiques avec précision. # PROCESSUS D'ANALYSE : # 1. Analysez le contexte fourni : {context} # 2. Utilisez la recherche web si la reponse n'existe pas dans le contexte # 3. Privilégiez les sources officielles et la jurisprudence récente # Question à traiter : {question} # """), # ("human", "{question}") # ]) # # Modified process_question function to better work with tuples # def process_question(question: str) -> Iterator[str]: # if question in question_cache: # response, docs = question_cache[question] # sources = [doc.metadata.get("source") for doc in docs] # sources = list(set([os.path.splitext(source)[0] for source in sources])) # yield response + "\n\n\nالمصادر المحتملة :\n" + "\n".join(sources) # return # relevant_docs = retriever(question) # # Reranking with cross-encoder # 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(key=lambda x: x[0], reverse=True) # reranked_docs = [d[2].page_content for d in scored_docs][:10] # 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 = [d[2].metadata['source'] for d in scored_docs][:10] # sources = list(set([os.path.splitext(source)[0] for source in sources])) # yield full_response + "\n\n\nالمصادر المحتملة :\n" + "\n".join(sources) # question_cache[question] = (full_response, relevant_docs) # except Exception as e: # yield f"Erreur lors du traitement : {str(e)}" # # Updated gradio_stream function for 'messages' format # def gradio_stream(question: str, chat_history: list) -> Iterator[list]: # # chat_history now contains the user message added by user_input # # Add a placeholder for the assistant's response # chat_history.append({"role": "assistant", "content": ""}) # try: # # Stream the response using the existing process_question generator # for partial_response in process_question(question): # # Update the content of the last message (the assistant's placeholder) # chat_history[-1]["content"] = partial_response # yield chat_history # Yield the entire updated history list # except Exception as e: # # Update the assistant's message with the error # chat_history[-1]["content"] = f"Erreur : {str(e)}" # yield chat_history # Yield the history with the error message # # Gradio interface # with gr.Blocks(css=css) as demo: # gr.Markdown("