| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385 |
- from flask import Flask, request, jsonify, send_from_directory
- from apscheduler.schedulers.background import BackgroundScheduler
- from sentence_transformers import CrossEncoder, SentenceTransformer
- from langchain_ollama import OllamaLLM
- from processing import full_update_pipeline, load_index_from_disk
- from agent_ai.auth import login_to_wiki
- import faiss
- import json
- import numpy as np
- import textwrap
- from sklearn.metrics.pairwise import cosine_similarity
- from tabulate import tabulate
- import os
- import threading
- from serpapi import GoogleSearch
- from werkzeug.utils import secure_filename
- import PyPDF2
- from PIL import Image
- from transformers import AutoTokenizer
- from ocr import run_ocr, generate_html , generate_word_level_html
- import tiktoken
- from VLM import ask_blip
- app = Flask(__name__)
- state = {
- "index": None,
- "all_chunks": [],
- "chunk_urls": [],
- "all_embeddings": None,
- "embedding_model": SentenceTransformer('all-MiniLM-L6-v2'),
- "cross_encoder": CrossEncoder("cross-encoder/ms-marco-MiniLM-L-6-v2"),
- "is_updating": False,
- "llm": OllamaLLM(model="llama3", temperature=0.3, top_p=0.9, max_tokens=600),
- "chat_history": []
-
- }
- tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
- def chunk_text(text, chunk_size=400, overlap=100):
- tokens = tokenizer.encode(text)
- chunks = []
- for i in range(0, len(tokens), chunk_size - overlap):
- chunk = tokens[i:i + chunk_size]
- chunks.append(tokenizer.decode(chunk))
- return chunks
- UPLOAD_FOLDER = "uploads"
- os.makedirs(UPLOAD_FOLDER, exist_ok=True) # Crée le dossier s'il n'existe pas
- app.config["UPLOAD_FOLDER"] = UPLOAD_FOLDER
- ALLOWED_EXTENSIONS = {'txt', 'pdf', 'png', 'jpg', 'jpeg'}
- def allowed_file(filename):
- return '.' in filename and filename.rsplit('.', 1)[1].lower() in ALLOWED_EXTENSIONS
- @app.route("/upload_file", methods=["POST"])
- def upload_file():
- if 'file' not in request.files:
- return jsonify({"error": "Aucun fichier détecté."}), 400
- file = request.files['file']
- if file.filename == '':
- return jsonify({"error": "Nom de fichier vide."}), 400
- if file and allowed_file(file.filename):
- filename = secure_filename(file.filename)
- filepath = os.path.join(app.config['UPLOAD_FOLDER'], filename)
- file.save(filepath)
- question = request.form.get("question", "")
- if not question:
- return jsonify({"error": "Veuillez poser une question liée au fichier."}), 400
- try:
- # ✅ 1. Exécuter l'OCR
- document, ocr_output = run_ocr(filepath)
- # ✅ 2. Sauvegarder le JSON OCR
- json_path = os.path.join(app.config['UPLOAD_FOLDER'], "ocr_output.json")
- with open(json_path, "w", encoding="utf-8") as fjson:
- json.dump(ocr_output, fjson, ensure_ascii=False, indent=2)
- # ✅ 3. Générer le HTML avec bounding boxes par mot
- html_path = os.path.join(app.config['UPLOAD_FOLDER'], "ocr_output_words.html")
- generate_word_level_html(ocr_output, output_html_path=html_path)
- # ✅ 4. Extraire le texte brut pour le traitement RAG
- text = ""
- for page in ocr_output['pages']:
- for block in page['blocks']:
- for line in block['lines']:
- for word in line['words']:
- text += word['value'] + " "
- text += "\n"
- text += "\n"
- text += "\n"
- except Exception as e:
- return jsonify({"error": f"OCR échoué : {str(e)}"}), 500
- if text.strip():
- chunks = chunk_text(text)
- embeddings = state["embedding_model"].encode(chunks)
- question_embedding = state["embedding_model"].encode([question])[0]
- similarities = cosine_similarity([question_embedding], embeddings)[0]
- top_indices = np.argsort(similarities)[::-1][:5]
- top_chunks = [chunks[i] for i in top_indices]
- top_context = "\n\n".join(top_chunks)
- prompt = (
- "Réponds en français à la question suivante en utilisant uniquement le contenu fourni. "
- "Si la réponse est absente, dis-le clairement.\n\n"
- f"{top_context}\n\nQuestion : {question}"
- )
- answer = state["llm"](prompt).strip()
- return jsonify({
- "answer": answer,
- "html_file": f"uploads/ocr_output_words.html"
- })
- else:
- try:
- answer = ask_blip(filepath, question)
- return jsonify({"answer": answer})
- except Exception as e:
- return jsonify({"error": f"LLaVA échoué : {str(e)}"}), 500
- return jsonify({"error": "Format de fichier non supporté."}), 400
- @app.route("/delete_file", methods=["POST"])
- def delete_file():
- data = request.get_json()
- filename = data.get("filename")
- if not filename:
- return jsonify({"error": "Nom de fichier manquant."}), 400
- filepath = os.path.join(app.config["UPLOAD_FOLDER"], secure_filename(filename))
- if os.path.exists(filepath):
- os.remove(filepath)
- return jsonify({"message": "Fichier supprimé avec succès."})
- else:
- return jsonify({"error": "Fichier introuvable."}), 404
- def load_or_update_index():
- print("🚀 App starting up...")
- index, texts, urls = load_index_from_disk()
- if index:
- state["index"] = index
- state["all_chunks"] = texts
- state["chunk_urls"] = urls
- print("✅ Index loaded from disk.")
- else:
- print("❌ No index found. Launching full update in background...")
- threading.Thread(target=run_full_update, daemon=True).start()
- def run_full_update():
- state["is_updating"] = True
- try:
- index, texts, urls = full_update_pipeline()
- if index:
- state["index"] = index
- state["all_chunks"] = texts
- state["chunk_urls"] = urls
- state["all_embeddings"] = state["embedding_model"].encode(texts)
- print("✅ Background update finished.")
- except Exception as e:
- print(f"❌ Failed to update index: {e}")
- finally:
- state["is_updating"] = False
- @app.route('/')
- def index():
- return send_from_directory(os.getcwd(), 'index.html')
- @app.route("/login", methods=["POST"])
- def login():
- data = request.json
- username = data.get("username")
- password = data.get("password")
- if not username or not password:
- return jsonify({"error": "username and password required"}), 400
- if not login_to_wiki(username, password):
- return jsonify({"error": "Authentication failed"}), 401
- return jsonify({"message": "Login successful"})
- @app.route("/chat_history", methods=["GET"])
- def chat_history():
- return jsonify(state["chat_history"])
- @app.route("/qa_chain", methods=["POST"])
- def qa_chain_route():
- if state.get("llm") is None:
- return jsonify({"error": "Le modèle LLM n'est pas initialisé."}), 500
- if state["index"] is None or not state["all_chunks"]:
- if state["is_updating"]:
- return jsonify({"error": "Le système est encore en train d’indexer les documents. Réessayez dans quelques instants."}), 503
- else:
- return jsonify({"error": "L’index est vide. Veuillez attendre la mise à jour ou réinitialiser le système."}), 500
- data = request.get_json()
- question = data.get("question", "")
- if not question:
- return jsonify({"error": "Question vide."}), 400
- # ÉTAPE 1 : Tentative de réponse avec le Wiki local
- try:
- query_embedding = state["embedding_model"].encode([question])[0]
- if state["all_embeddings"] is None:
- state["all_embeddings"] = state["embedding_model"].encode(state["all_chunks"])
- similarities = cosine_similarity([query_embedding], state["all_embeddings"])[0]
- top_indices = np.argsort(similarities)[::-1][:1500]
- rerank_data = [(question, state["all_chunks"][i]) for i in top_indices]
- rerank_scores = state["cross_encoder"].predict(rerank_data)
- reranked = sorted(
- zip([state["all_chunks"][i] for i in top_indices], top_indices, rerank_scores),
- key=lambda x: x[2], reverse=True
- )
-
- # Sauvegarde du contexte dans un fichier de cache
- enc = tiktoken.encoding_for_model("gpt-3.5-turbo")
- cache_data = []
- for rank, (chunk, idx, score) in enumerate(reranked[:1500], start=1):
- url = state["chunk_urls"][idx] if idx < len(state["chunk_urls"]) else "URL inconnue"
- token_count = len(enc.encode(chunk))
- cache_data.append({
- "rank": rank,
- "url": url,
- "chunk": chunk,
- "tokens": token_count
- })
- with open("cache_context.json", "w", encoding="utf-8") as f:
- json.dump(cache_data, f, indent=2, ensure_ascii=False)
- # Affichage en terminal
- table = []
- top_chunks = []
- for rank, (chunk, idx, score) in enumerate(reranked, start=1):
- url = state["chunk_urls"][idx] if idx < len(state["chunk_urls"]) else "URL inconnue"
- wrapped_chunk = "\n".join(textwrap.wrap(chunk.strip(), width=80))[:1000] + "..."
- wrapped_url = "\n".join(textwrap.wrap(url, width=80))
- table.append([rank, round(score, 4), wrapped_chunk, wrapped_url])
- top_chunks.append(chunk)
- print("\n================== RÉSULTATS CLASSÉS PAR PERTINENCE ==================")
- print(tabulate(table, headers=["Rank", "Score", "Chunk", "URL"], tablefmt="fancy_grid"))
- print("=======================================================================\n")
- max_chars = 15000
- top_chunks = []
- current_length = 0
- for chunk, _, _ in reranked:
- if current_length + len(chunk) > max_chars:
- break
- top_chunks.append(chunk)
- current_length += len(chunk)
- top_context = "\n\n".join(top_chunks)
- prompt = (
- "Réponds uniquement à la question en utilisant les informations contenues dans le contexte ci-dessous. "
- "Ta réponse doit être factuelle, concise et précise. "
- "N’ajoute aucune information extérieure. "
- "Si la réponse n’est pas présente dans le contexte, dis simplement : 'Information non disponible.'\n\n"
- f"Contexte :\n{top_context}\n\n"
- f"Question : {question}"
- )
- local_answer = state["llm"](prompt).strip()
- print("\n====================== PROMPT UTILISÉ ======================")
- print(f"CONTEXTE :\n{top_context}\nQUESTION :\n{question}\nRÉPONSE :\n{local_answer}")
- print("============================================================\n")
- # Vérifie si la réponse est valable
- phrases_non_reponse = [
- "n'est pas disponible",
- "je ne sais pas",
- "je ne trouve pas",
- "aucune information",
- "information non disponible",
- "pas d'information",
- "je ne suis pas en mesure",
- "le contexte ne contient pas",
- "https://localhost:8080"
- ]
- if local_answer and not any(phrase in local_answer.lower() for phrase in phrases_non_reponse):
- source_urls = [state["chunk_urls"][idx] for _, idx, _ in reranked[:5]]
- state["chat_history"].append({
- "question": question,
- "answer": local_answer,
- "sources": list(set(source_urls))
- })
- return jsonify({
- "answer": local_answer,
- "sources": list(set(source_urls))
- })
- except Exception as e:
- print(f"[Erreur Wiki] : {e}")
- # ÉTAPE 2 : Fallback vers recherche web (SerpAPI)
- try:
- params = {
- "engine": "google",
- "q": question,
- "api_key": "1b0f09bd610259b457b08c5fad06ded73a9e108e6bcc932e02ac88b2f3840171",
- "gl": "fr",
- "hl": "fr"
- }
- search = GoogleSearch(params)
- results = search.get_dict()
- answer = "Je n'ai pas trouvé de réponse précise."
- if "answer_box" in results and "answer" in results["answer_box"]:
- answer = results["answer_box"]["answer"]
- elif "organic_results" in results and results["organic_results"]:
- answer = results["organic_results"][0].get("snippet", answer)
- # Résumer ou filtrer la réponse via le LLM
- if state.get("llm"):
- prompt = (
- "Réponds uniquement à la question de manière directe et concise.\n"
- "Sans reformuler la question, sans phrase d’introduction, sans politesse et sans contenu promotionnel. "
- "Utilise uniquement les faits du texte ci-dessous. "
- "Si l’information n’est pas présente, réponds : 'Information non disponible.'\n\n"
- f"Texte : {answer}\n\n"
- f"Question : {question}"
- )
- try:
- answer = state["llm"](prompt).strip()
- except Exception as llm_error:
- print("[Erreur LLM sur reformulation] :", llm_error)
- # Historique
- state["chat_history"].append({
- "question": question,
- "answer": answer,
- "sources": ["Google Search"]
- })
- return jsonify({
- "answer": answer,
- "sources": ["Google Search"]
- })
- except Exception as e:
- return jsonify({"error": f"Erreur SerpAPI : {str(e)}"}), 500
- # Scheduler
- scheduler = BackgroundScheduler()
- scheduler.add_job(run_full_update, 'interval', minutes=30)
- print("🔁 Scheduler activé pour mises à jour toutes les 30 min")
-
- if __name__ == "__main__":
-
- load_or_update_index()
- scheduler.start()
- app.run(debug=True, use_reloader=False)
|