speed it
Browse files
app.py
CHANGED
@@ -1,309 +1,548 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
from langchain_mistralai.chat_models import ChatMistralAI
|
3 |
from langchain.prompts import ChatPromptTemplate
|
4 |
import os
|
5 |
from pathlib import Path
|
6 |
-
from typing import List, Dict, Optional
|
7 |
import json
|
8 |
import faiss
|
9 |
import numpy as np
|
10 |
from langchain.schema import Document
|
11 |
-
from sentence_transformers import SentenceTransformer
|
12 |
import pickle
|
13 |
import re
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
-
|
16 |
-
|
17 |
-
class RAGLoader:
|
18 |
def __init__(self,
|
19 |
docs_folder: str = "./docs",
|
20 |
splits_folder: str = "./splits",
|
21 |
-
index_folder: str = "./index"
|
22 |
-
|
23 |
-
"""
|
24 |
-
Initialise le RAG Loader
|
25 |
-
|
26 |
-
Args:
|
27 |
-
docs_folder: Dossier contenant les documents sources
|
28 |
-
splits_folder: Dossier où seront stockés les morceaux de texte
|
29 |
-
index_folder: Dossier où sera stocké l'index FAISS
|
30 |
-
model_name: Nom du modèle SentenceTransformer à utiliser
|
31 |
-
"""
|
32 |
self.docs_folder = Path(docs_folder)
|
33 |
self.splits_folder = Path(splits_folder)
|
34 |
self.index_folder = Path(index_folder)
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
# Chemins des fichiers
|
42 |
self.splits_path = self.splits_folder / "splits.json"
|
43 |
self.index_path = self.index_folder / "faiss.index"
|
44 |
self.documents_path = self.index_folder / "documents.pkl"
|
45 |
-
|
46 |
-
#
|
47 |
-
# self.model = None
|
48 |
self.index = None
|
49 |
self.indexed_documents = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
50 |
|
51 |
-
def
|
52 |
-
token = os.environ.get('HUGGINGFACE_TOKEN')
|
53 |
-
API_URL = "https://api-inference.huggingface.co/models/intfloat/multilingual-e5-large"
|
54 |
-
headers = {"Authorization": "Bearer {token}"}
|
55 |
-
response = requests.post(API_URL, headers=headers, json=payload)
|
56 |
-
return response.json()
|
57 |
-
|
58 |
-
def load_and_split_texts(self) -> List[Document]:
|
59 |
-
"""
|
60 |
-
Charge les textes du dossier docs, les découpe en morceaux et les sauvegarde
|
61 |
-
dans un fichier JSON unique.
|
62 |
-
|
63 |
-
Returns:
|
64 |
-
Liste de Documents contenant les morceaux de texte et leurs métadonnées
|
65 |
-
"""
|
66 |
-
documents = []
|
67 |
-
|
68 |
-
# Vérifier d'abord si les splits existent déjà
|
69 |
if self._splits_exist():
|
70 |
-
print("Chargement des splits existants...")
|
71 |
return self._load_existing_splits()
|
72 |
-
|
73 |
-
|
74 |
-
|
|
|
75 |
for file_path in self.docs_folder.glob("*.txt"):
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
# Créer un Document pour chaque morceau
|
84 |
-
for i, chunk in enumerate(chunks):
|
85 |
-
doc = Document(
|
86 |
-
page_content=chunk,
|
87 |
-
metadata={
|
88 |
-
'source': file_path.name,
|
89 |
-
'chunk_id': i,
|
90 |
-
'total_chunks': len(chunks)
|
91 |
-
}
|
92 |
-
)
|
93 |
-
documents.append(doc)
|
94 |
-
|
95 |
-
# Sauvegarder tous les splits dans un seul fichier JSON
|
96 |
self._save_splits(documents)
|
97 |
-
|
98 |
-
print(f"Nombre total de morceaux créés: {len(documents)}")
|
99 |
return documents
|
100 |
-
|
101 |
-
def
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
|
|
|
|
114 |
]
|
115 |
-
}
|
116 |
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
with open(self.splits_path, 'r', encoding='utf-8') as f:
|
123 |
-
splits_data = json.load(f)
|
124 |
-
|
125 |
-
documents = [
|
126 |
-
Document(
|
127 |
-
page_content=split['text'],
|
128 |
-
metadata=split['metadata']
|
129 |
-
)
|
130 |
-
for split in splits_data['splits']
|
131 |
-
]
|
132 |
-
|
133 |
-
print(f"Nombre de splits chargés: {len(documents)}")
|
134 |
-
return documents
|
135 |
-
|
136 |
-
def load_index(self) -> bool:
|
137 |
-
"""
|
138 |
-
Charge l'index FAISS et les documents associés s'ils existent
|
139 |
-
|
140 |
-
Returns:
|
141 |
-
bool: True si l'index a été chargé, False sinon
|
142 |
-
"""
|
143 |
-
if not self._index_exists():
|
144 |
-
print("Aucun index trouvé.")
|
145 |
-
return False
|
146 |
-
|
147 |
-
print("Chargement de l'index existant...")
|
148 |
-
try:
|
149 |
-
# Charger l'index FAISS
|
150 |
-
self.index = faiss.read_index(str(self.index_path))
|
151 |
-
|
152 |
-
# Charger les documents associés
|
153 |
-
with open(self.documents_path, 'rb') as f:
|
154 |
-
self.indexed_documents = pickle.load(f)
|
155 |
-
|
156 |
-
print(f"Index chargé avec {self.index.ntotal} vecteurs")
|
157 |
-
return True
|
158 |
-
|
159 |
-
except Exception as e:
|
160 |
-
print(f"Erreur lors du chargement de l'index: {e}")
|
161 |
return False
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
print("Aucun document à indexer.")
|
186 |
-
return False
|
187 |
-
|
188 |
-
print("Création des embeddings...")
|
189 |
-
texts = [doc.page_content for doc in documents]
|
190 |
-
embeddings = self.encode(texts)
|
191 |
-
|
192 |
-
# Initialiser l'index FAISS
|
193 |
-
dimension = embeddings.shape[1]
|
194 |
-
self.index = faiss.IndexFlatL2(dimension)
|
195 |
-
|
196 |
-
# Ajouter les vecteurs à l'index
|
197 |
-
self.index.add(np.array(embeddings).astype('float32'))
|
198 |
-
|
199 |
-
# Sauvegarder l'index
|
200 |
-
print("Sauvegarde de l'index...")
|
201 |
-
faiss.write_index(self.index, str(self.index_path))
|
202 |
-
|
203 |
-
# Sauvegarder les documents associés
|
204 |
-
self.indexed_documents = documents
|
205 |
-
with open(self.documents_path, 'wb') as f:
|
206 |
-
pickle.dump(documents, f)
|
207 |
-
|
208 |
-
print(f"Index créé avec succès : {self.index.ntotal} vecteurs")
|
209 |
-
return True
|
210 |
-
|
211 |
-
except Exception as e:
|
212 |
-
print(f"Erreur lors de la création de l'index: {e}")
|
213 |
-
return False
|
214 |
-
|
215 |
-
def _index_exists(self) -> bool:
|
216 |
-
"""Vérifie si l'index et les documents associés existent"""
|
217 |
-
return self.index_path.exists() and self.documents_path.exists()
|
218 |
|
219 |
def get_retriever(self, k: int = 10):
|
220 |
-
"""
|
221 |
-
Crée un retriever pour l'utilisation avec LangChain
|
222 |
-
|
223 |
-
Args:
|
224 |
-
k: Nombre de documents similaires à retourner
|
225 |
-
|
226 |
-
Returns:
|
227 |
-
Callable: Fonction de recherche compatible avec LangChain
|
228 |
-
"""
|
229 |
if self.index is None:
|
230 |
if not self.load_index():
|
231 |
if not self.create_index():
|
232 |
-
raise ValueError("
|
233 |
|
234 |
-
|
235 |
-
|
|
|
|
|
|
|
236 |
|
237 |
-
|
238 |
-
|
239 |
-
query_embedding = self.encode([query])[0]
|
240 |
-
|
241 |
-
# Rechercher les documents similaires
|
242 |
distances, indices = self.index.search(
|
243 |
np.array([query_embedding]).astype('float32'),
|
244 |
k
|
245 |
)
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
if idx != -1
|
251 |
-
|
252 |
-
|
|
|
|
|
253 |
return results
|
254 |
-
|
255 |
return retriever_function
|
256 |
|
257 |
-
# Initialize
|
258 |
-
llm = ChatMistralAI(
|
259 |
-
|
260 |
-
|
261 |
-
|
262 |
-
|
263 |
-
("system", """أنت مساعد مفيد يجيب على الأسئلة باللغة العربية باستخدام المعلومات المقدمة.
|
264 |
-
استخدم المعلومات التالية للإجابة على السؤال:
|
265 |
|
266 |
-
|
|
|
267 |
|
268 |
-
|
269 |
-
|
270 |
-
("human", "{question}")
|
271 |
-
])
|
272 |
|
273 |
def process_question(question: str) -> tuple[str, str]:
|
274 |
-
|
275 |
-
|
276 |
-
|
|
|
277 |
relevant_docs = retriever(question)
|
278 |
context = "\n".join([doc.page_content for doc in relevant_docs])
|
279 |
-
|
280 |
prompt = prompt_template.format_messages(
|
281 |
context=context,
|
282 |
question=question
|
283 |
)
|
284 |
-
|
285 |
response = llm(prompt)
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
-
|
294 |
-
# Custom CSS for right-aligned text in textboxes
|
295 |
-
custom_css = """
|
296 |
-
.rtl-text {
|
297 |
-
text-align: right !important;
|
298 |
-
direction: rtl !important;
|
299 |
-
}
|
300 |
-
.rtl-text textarea {
|
301 |
-
text-align: right !important;
|
302 |
-
direction: rtl !important;
|
303 |
-
}
|
304 |
-
"""
|
305 |
-
|
306 |
-
# Define the Gradio interface
|
307 |
with gr.Blocks(css=custom_css) as iface:
|
308 |
with gr.Column():
|
309 |
input_text = gr.Textbox(
|
@@ -312,27 +551,33 @@ with gr.Blocks(css=custom_css) as iface:
|
|
312 |
lines=2,
|
313 |
elem_classes="rtl-text"
|
314 |
)
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
321 |
-
|
322 |
-
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
submit_btn = gr.Button("إرسال")
|
329 |
-
|
330 |
submit_btn.click(
|
331 |
-
fn=
|
332 |
inputs=input_text,
|
333 |
-
outputs=[answer_box, context_box]
|
|
|
334 |
)
|
335 |
|
336 |
-
# Launch
|
337 |
if __name__ == "__main__":
|
338 |
-
iface.
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# import gradio as gr
|
2 |
+
# from langchain_mistralai.chat_models import ChatMistralAI
|
3 |
+
# from langchain.prompts import ChatPromptTemplate
|
4 |
+
# import os
|
5 |
+
# from pathlib import Path
|
6 |
+
# from typing import List, Dict, Optional
|
7 |
+
# import json
|
8 |
+
# import faiss
|
9 |
+
# import numpy as np
|
10 |
+
# from langchain.schema import Document
|
11 |
+
# from sentence_transformers import SentenceTransformer
|
12 |
+
# import pickle
|
13 |
+
# import re
|
14 |
+
|
15 |
+
# os.environ.get('HUGGINGFACE_TOKEN')
|
16 |
+
|
17 |
+
# class RAGLoader:
|
18 |
+
# def __init__(self,
|
19 |
+
# docs_folder: str = "./docs",
|
20 |
+
# splits_folder: str = "./splits",
|
21 |
+
# index_folder: str = "./index",):
|
22 |
+
# # model_name: str = "intfloat/multilingual-e5-large")
|
23 |
+
# """
|
24 |
+
# Initialise le RAG Loader
|
25 |
+
|
26 |
+
# Args:
|
27 |
+
# docs_folder: Dossier contenant les documents sources
|
28 |
+
# splits_folder: Dossier où seront stockés les morceaux de texte
|
29 |
+
# index_folder: Dossier où sera stocké l'index FAISS
|
30 |
+
# model_name: Nom du modèle SentenceTransformer à utiliser
|
31 |
+
# """
|
32 |
+
# self.docs_folder = Path(docs_folder)
|
33 |
+
# self.splits_folder = Path(splits_folder)
|
34 |
+
# self.index_folder = Path(index_folder)
|
35 |
+
# # self.model_name = model_name
|
36 |
+
|
37 |
+
# # Créer les dossiers s'ils n'existent pas
|
38 |
+
# self.splits_folder.mkdir(parents=True, exist_ok=True)
|
39 |
+
# self.index_folder.mkdir(parents=True, exist_ok=True)
|
40 |
+
|
41 |
+
# # Chemins des fichiers
|
42 |
+
# self.splits_path = self.splits_folder / "splits.json"
|
43 |
+
# self.index_path = self.index_folder / "faiss.index"
|
44 |
+
# self.documents_path = self.index_folder / "documents.pkl"
|
45 |
+
|
46 |
+
# # Initialiser le modèle
|
47 |
+
# # self.model = None
|
48 |
+
# self.index = None
|
49 |
+
# self.indexed_documents = None
|
50 |
+
|
51 |
+
# def encode(self,payload):
|
52 |
+
# token = os.environ.get('HUGGINGFACE_TOKEN')
|
53 |
+
# API_URL = "https://api-inference.huggingface.co/models/intfloat/multilingual-e5-large"
|
54 |
+
# headers = {"Authorization": "Bearer {token}"}
|
55 |
+
# response = requests.post(API_URL, headers=headers, json=payload)
|
56 |
+
# return response.json()
|
57 |
+
|
58 |
+
# def load_and_split_texts(self) -> List[Document]:
|
59 |
+
# """
|
60 |
+
# Charge les textes du dossier docs, les découpe en morceaux et les sauvegarde
|
61 |
+
# dans un fichier JSON unique.
|
62 |
+
|
63 |
+
# Returns:
|
64 |
+
# Liste de Documents contenant les morceaux de texte et leurs métadonnées
|
65 |
+
# """
|
66 |
+
# documents = []
|
67 |
+
|
68 |
+
# # Vérifier d'abord si les splits existent déjà
|
69 |
+
# if self._splits_exist():
|
70 |
+
# print("Chargement des splits existants...")
|
71 |
+
# return self._load_existing_splits()
|
72 |
+
|
73 |
+
# print("Création de nouveaux splits...")
|
74 |
+
# # Parcourir tous les fichiers du dossier docs
|
75 |
+
# for file_path in self.docs_folder.glob("*.txt"):
|
76 |
+
# with open(file_path, 'r', encoding='utf-8') as file:
|
77 |
+
# text = file.read()
|
78 |
+
|
79 |
+
# # Découper le texte en phrases
|
80 |
+
# # chunks = [chunk.strip() for chunk in text.split('.') if chunk.strip()]
|
81 |
+
# chunks = [s.strip() for s in re.split(r'(?<=[.!?])\s+', text) if s.strip()]
|
82 |
+
|
83 |
+
# # Créer un Document pour chaque morceau
|
84 |
+
# for i, chunk in enumerate(chunks):
|
85 |
+
# doc = Document(
|
86 |
+
# page_content=chunk,
|
87 |
+
# metadata={
|
88 |
+
# 'source': file_path.name,
|
89 |
+
# 'chunk_id': i,
|
90 |
+
# 'total_chunks': len(chunks)
|
91 |
+
# }
|
92 |
+
# )
|
93 |
+
# documents.append(doc)
|
94 |
+
|
95 |
+
# # Sauvegarder tous les splits dans un seul fichier JSON
|
96 |
+
# self._save_splits(documents)
|
97 |
+
|
98 |
+
# print(f"Nombre total de morceaux créés: {len(documents)}")
|
99 |
+
# return documents
|
100 |
+
|
101 |
+
# def _splits_exist(self) -> bool:
|
102 |
+
# """Vérifie si le fichier de splits existe"""
|
103 |
+
# return self.splits_path.exists()
|
104 |
+
|
105 |
+
# def _save_splits(self, documents: List[Document]):
|
106 |
+
# """Sauvegarde tous les documents découpés dans un seul fichier JSON"""
|
107 |
+
# splits_data = {
|
108 |
+
# 'splits': [
|
109 |
+
# {
|
110 |
+
# 'text': doc.page_content,
|
111 |
+
# 'metadata': doc.metadata
|
112 |
+
# }
|
113 |
+
# for doc in documents
|
114 |
+
# ]
|
115 |
+
# }
|
116 |
+
|
117 |
+
# with open(self.splits_path, 'w', encoding='utf-8') as f:
|
118 |
+
# json.dump(splits_data, f, ensure_ascii=False, indent=2)
|
119 |
+
|
120 |
+
# def _load_existing_splits(self) -> List[Document]:
|
121 |
+
# """Charge les splits depuis le fichier JSON unique"""
|
122 |
+
# with open(self.splits_path, 'r', encoding='utf-8') as f:
|
123 |
+
# splits_data = json.load(f)
|
124 |
+
|
125 |
+
# documents = [
|
126 |
+
# Document(
|
127 |
+
# page_content=split['text'],
|
128 |
+
# metadata=split['metadata']
|
129 |
+
# )
|
130 |
+
# for split in splits_data['splits']
|
131 |
+
# ]
|
132 |
+
|
133 |
+
# print(f"Nombre de splits chargés: {len(documents)}")
|
134 |
+
# return documents
|
135 |
+
|
136 |
+
# def load_index(self) -> bool:
|
137 |
+
# """
|
138 |
+
# Charge l'index FAISS et les documents associés s'ils existent
|
139 |
+
|
140 |
+
# Returns:
|
141 |
+
# bool: True si l'index a été chargé, False sinon
|
142 |
+
# """
|
143 |
+
# if not self._index_exists():
|
144 |
+
# print("Aucun index trouvé.")
|
145 |
+
# return False
|
146 |
+
|
147 |
+
# print("Chargement de l'index existant...")
|
148 |
+
# try:
|
149 |
+
# # Charger l'index FAISS
|
150 |
+
# self.index = faiss.read_index(str(self.index_path))
|
151 |
+
|
152 |
+
# # Charger les documents associés
|
153 |
+
# with open(self.documents_path, 'rb') as f:
|
154 |
+
# self.indexed_documents = pickle.load(f)
|
155 |
+
|
156 |
+
# print(f"Index chargé avec {self.index.ntotal} vecteurs")
|
157 |
+
# return True
|
158 |
+
|
159 |
+
# except Exception as e:
|
160 |
+
# print(f"Erreur lors du chargement de l'index: {e}")
|
161 |
+
# return False
|
162 |
+
|
163 |
+
# def create_index(self, documents: Optional[List[Document]] = None) -> bool:
|
164 |
+
# """
|
165 |
+
# Crée un nouvel index FAISS à partir des documents.
|
166 |
+
# Si aucun document n'est fourni, charge les documents depuis le fichier JSON.
|
167 |
+
|
168 |
+
# Args:
|
169 |
+
# documents: Liste optionnelle de Documents à indexer
|
170 |
+
|
171 |
+
# Returns:
|
172 |
+
# bool: True si l'index a été créé avec succès, False sinon
|
173 |
+
# """
|
174 |
+
# try:
|
175 |
+
# # # Initialiser le modèle si nécessaire
|
176 |
+
# # if self.model is None:
|
177 |
+
# # print("Chargement du modèle...")
|
178 |
+
# # self.model = SentenceTransformer(self.model_name)
|
179 |
+
|
180 |
+
# # Charger les documents si non fournis
|
181 |
+
# if documents is None:
|
182 |
+
# documents = self.load_and_split_texts()
|
183 |
+
|
184 |
+
# if not documents:
|
185 |
+
# print("Aucun document à indexer.")
|
186 |
+
# return False
|
187 |
+
|
188 |
+
# print("Création des embeddings...")
|
189 |
+
# texts = [doc.page_content for doc in documents]
|
190 |
+
# embeddings = self.encode(texts)
|
191 |
+
|
192 |
+
# # Initialiser l'index FAISS
|
193 |
+
# dimension = embeddings.shape[1]
|
194 |
+
# self.index = faiss.IndexFlatL2(dimension)
|
195 |
+
|
196 |
+
# # Ajouter les vecteurs à l'index
|
197 |
+
# self.index.add(np.array(embeddings).astype('float32'))
|
198 |
+
|
199 |
+
# # Sauvegarder l'index
|
200 |
+
# print("Sauvegarde de l'index...")
|
201 |
+
# faiss.write_index(self.index, str(self.index_path))
|
202 |
+
|
203 |
+
# # Sauvegarder les documents associés
|
204 |
+
# self.indexed_documents = documents
|
205 |
+
# with open(self.documents_path, 'wb') as f:
|
206 |
+
# pickle.dump(documents, f)
|
207 |
+
|
208 |
+
# print(f"Index créé avec succès : {self.index.ntotal} vecteurs")
|
209 |
+
# return True
|
210 |
+
|
211 |
+
# except Exception as e:
|
212 |
+
# print(f"Erreur lors de la création de l'index: {e}")
|
213 |
+
# return False
|
214 |
+
|
215 |
+
# def _index_exists(self) -> bool:
|
216 |
+
# """Vérifie si l'index et les documents associés existent"""
|
217 |
+
# return self.index_path.exists() and self.documents_path.exists()
|
218 |
+
|
219 |
+
# def get_retriever(self, k: int = 10):
|
220 |
+
# """
|
221 |
+
# Crée un retriever pour l'utilisation avec LangChain
|
222 |
+
|
223 |
+
# Args:
|
224 |
+
# k: Nombre de documents similaires à retourner
|
225 |
+
|
226 |
+
# Returns:
|
227 |
+
# Callable: Fonction de recherche compatible avec LangChain
|
228 |
+
# """
|
229 |
+
# if self.index is None:
|
230 |
+
# if not self.load_index():
|
231 |
+
# if not self.create_index():
|
232 |
+
# raise ValueError("Impossible de charger ou créer l'index")
|
233 |
+
|
234 |
+
# # if self.model is None:
|
235 |
+
# # self.model = SentenceTransformer(self.model_name)
|
236 |
+
|
237 |
+
# def retriever_function(query: str) -> List[Document]:
|
238 |
+
# # Créer l'embedding de la requête
|
239 |
+
# query_embedding = self.encode([query])[0]
|
240 |
+
|
241 |
+
# # Rechercher les documents similaires
|
242 |
+
# distances, indices = self.index.search(
|
243 |
+
# np.array([query_embedding]).astype('float32'),
|
244 |
+
# k
|
245 |
+
# )
|
246 |
+
|
247 |
+
# # Retourner les documents trouvés
|
248 |
+
# results = []
|
249 |
+
# for idx in indices[0]:
|
250 |
+
# if idx != -1: # FAISS retourne -1 pour les résultats invalides
|
251 |
+
# results.append(self.indexed_documents[idx])
|
252 |
+
|
253 |
+
# return results
|
254 |
+
|
255 |
+
# return retriever_function
|
256 |
+
|
257 |
+
# # Initialize the RAG system
|
258 |
+
# llm = ChatMistralAI(model="mistral-large-latest", mistral_api_key="QK0ZZpSxQbCEVgOLtI6FARQVmBYc6WGP")
|
259 |
+
# rag_loader = RAGLoader()
|
260 |
+
# retriever = rag_loader.get_retriever(k=10)
|
261 |
+
|
262 |
+
# prompt_template = ChatPromptTemplate.from_messages([
|
263 |
+
# ("system", """أنت مساعد مفيد يجيب على الأسئلة باللغة العربية باستخدام المعلومات المقدمة.
|
264 |
+
# استخدم المعلومات التالية للإجابة على السؤال:
|
265 |
+
|
266 |
+
# {context}
|
267 |
+
|
268 |
+
# إذا لم تكن المعلومات كافية للإجابة على السؤال بشكل كامل، قم بتوضيح ذلك.
|
269 |
+
# أجب بشكل موجز ودقيق."""),
|
270 |
+
# ("human", "{question}")
|
271 |
+
# ])
|
272 |
+
|
273 |
+
# def process_question(question: str) -> tuple[str, str]:
|
274 |
+
# """
|
275 |
+
# Process a question and return both the answer and the relevant context
|
276 |
+
# """
|
277 |
+
# relevant_docs = retriever(question)
|
278 |
+
# context = "\n".join([doc.page_content for doc in relevant_docs])
|
279 |
+
|
280 |
+
# prompt = prompt_template.format_messages(
|
281 |
+
# context=context,
|
282 |
+
# question=question
|
283 |
+
# )
|
284 |
+
|
285 |
+
# response = llm(prompt)
|
286 |
+
# return response.content, context
|
287 |
+
|
288 |
+
# def gradio_interface(question: str) -> tuple[str, str]:
|
289 |
+
# """
|
290 |
+
# Gradio interface function that returns both answer and context as a tuple
|
291 |
+
# """
|
292 |
+
# return process_question(question)
|
293 |
+
|
294 |
+
# # Custom CSS for right-aligned text in textboxes
|
295 |
+
# custom_css = """
|
296 |
+
# .rtl-text {
|
297 |
+
# text-align: right !important;
|
298 |
+
# direction: rtl !important;
|
299 |
+
# }
|
300 |
+
# .rtl-text textarea {
|
301 |
+
# text-align: right !important;
|
302 |
+
# direction: rtl !important;
|
303 |
+
# }
|
304 |
+
# """
|
305 |
+
|
306 |
+
# # Define the Gradio interface
|
307 |
+
# with gr.Blocks(css=custom_css) as iface:
|
308 |
+
# with gr.Column():
|
309 |
+
# input_text = gr.Textbox(
|
310 |
+
# label="السؤال",
|
311 |
+
# placeholder="اكتب سؤالك هنا...",
|
312 |
+
# lines=2,
|
313 |
+
# elem_classes="rtl-text"
|
314 |
+
# )
|
315 |
+
|
316 |
+
# answer_box = gr.Textbox(
|
317 |
+
# label="الإجابة",
|
318 |
+
# lines=4,
|
319 |
+
# elem_classes="rtl-text"
|
320 |
+
# )
|
321 |
+
|
322 |
+
# context_box = gr.Textbox(
|
323 |
+
# label="السياق المستخدم",
|
324 |
+
# lines=8,
|
325 |
+
# elem_classes="rtl-text"
|
326 |
+
# )
|
327 |
+
|
328 |
+
# submit_btn = gr.Button("إرسال")
|
329 |
+
|
330 |
+
# submit_btn.click(
|
331 |
+
# fn=gradio_interface,
|
332 |
+
# inputs=input_text,
|
333 |
+
# outputs=[answer_box, context_box]
|
334 |
+
# )
|
335 |
+
|
336 |
+
# # Launch the interface
|
337 |
+
# if __name__ == "__main__":
|
338 |
+
# iface.launch(share=True)
|
339 |
+
|
340 |
+
|
341 |
import gradio as gr
|
342 |
from langchain_mistralai.chat_models import ChatMistralAI
|
343 |
from langchain.prompts import ChatPromptTemplate
|
344 |
import os
|
345 |
from pathlib import Path
|
|
|
346 |
import json
|
347 |
import faiss
|
348 |
import numpy as np
|
349 |
from langchain.schema import Document
|
|
|
350 |
import pickle
|
351 |
import re
|
352 |
+
import requests
|
353 |
+
from functools import lru_cache
|
354 |
+
import torch
|
355 |
+
from sentence_transformers import SentenceTransformer
|
356 |
+
import threading
|
357 |
+
from queue import Queue
|
358 |
+
import concurrent.futures
|
359 |
|
360 |
+
class OptimizedRAGLoader:
|
|
|
|
|
361 |
def __init__(self,
|
362 |
docs_folder: str = "./docs",
|
363 |
splits_folder: str = "./splits",
|
364 |
+
index_folder: str = "./index"):
|
365 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
366 |
self.docs_folder = Path(docs_folder)
|
367 |
self.splits_folder = Path(splits_folder)
|
368 |
self.index_folder = Path(index_folder)
|
369 |
+
|
370 |
+
# Create folders if they don't exist
|
371 |
+
for folder in [self.splits_folder, self.index_folder]:
|
372 |
+
folder.mkdir(parents=True, exist_ok=True)
|
373 |
+
|
374 |
+
# File paths
|
|
|
375 |
self.splits_path = self.splits_folder / "splits.json"
|
376 |
self.index_path = self.index_folder / "faiss.index"
|
377 |
self.documents_path = self.index_folder / "documents.pkl"
|
378 |
+
|
379 |
+
# Initialize components
|
|
|
380 |
self.index = None
|
381 |
self.indexed_documents = None
|
382 |
+
|
383 |
+
# Initialize encoder model
|
384 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
385 |
+
self.encoder = SentenceTransformer("intfloat/multilingual-e5-large")
|
386 |
+
self.encoder.to(self.device)
|
387 |
+
|
388 |
+
# Initialize thread pool
|
389 |
+
self.executor = concurrent.futures.ThreadPoolExecutor(max_workers=4)
|
390 |
+
|
391 |
+
# Initialize response cache
|
392 |
+
self.response_cache = {}
|
393 |
+
|
394 |
+
@lru_cache(maxsize=1000)
|
395 |
+
def encode(self, text: str):
|
396 |
+
"""Cached encoding function"""
|
397 |
+
with torch.no_grad():
|
398 |
+
embeddings = self.encoder.encode(
|
399 |
+
text,
|
400 |
+
convert_to_numpy=True,
|
401 |
+
normalize_embeddings=True
|
402 |
+
)
|
403 |
+
return embeddings
|
404 |
+
|
405 |
+
def batch_encode(self, texts: list):
|
406 |
+
"""Batch encoding for multiple texts"""
|
407 |
+
with torch.no_grad():
|
408 |
+
embeddings = self.encoder.encode(
|
409 |
+
texts,
|
410 |
+
batch_size=32,
|
411 |
+
convert_to_numpy=True,
|
412 |
+
normalize_embeddings=True,
|
413 |
+
show_progress_bar=False
|
414 |
+
)
|
415 |
+
return embeddings
|
416 |
|
417 |
+
def load_and_split_texts(self):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
418 |
if self._splits_exist():
|
|
|
419 |
return self._load_existing_splits()
|
420 |
+
|
421 |
+
documents = []
|
422 |
+
futures = []
|
423 |
+
|
424 |
for file_path in self.docs_folder.glob("*.txt"):
|
425 |
+
future = self.executor.submit(self._process_file, file_path)
|
426 |
+
futures.append(future)
|
427 |
+
|
428 |
+
for future in concurrent.futures.as_completed(futures):
|
429 |
+
documents.extend(future.result())
|
430 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
431 |
self._save_splits(documents)
|
|
|
|
|
432 |
return documents
|
433 |
+
|
434 |
+
def _process_file(self, file_path):
|
435 |
+
with open(file_path, 'r', encoding='utf-8') as file:
|
436 |
+
text = file.read()
|
437 |
+
chunks = [s.strip() for s in re.split(r'(?<=[.!?])\s+', text) if s.strip()]
|
438 |
+
|
439 |
+
return [
|
440 |
+
Document(
|
441 |
+
page_content=chunk,
|
442 |
+
metadata={
|
443 |
+
'source': file_path.name,
|
444 |
+
'chunk_id': i,
|
445 |
+
'total_chunks': len(chunks)
|
446 |
+
}
|
447 |
+
)
|
448 |
+
for i, chunk in enumerate(chunks)
|
449 |
]
|
|
|
450 |
|
451 |
+
def create_index(self, documents=None):
|
452 |
+
if documents is None:
|
453 |
+
documents = self.load_and_split_texts()
|
454 |
+
|
455 |
+
if not documents:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
456 |
return False
|
457 |
+
|
458 |
+
texts = [doc.page_content for doc in documents]
|
459 |
+
embeddings = self.batch_encode(texts)
|
460 |
+
|
461 |
+
dimension = embeddings.shape[1]
|
462 |
+
self.index = faiss.IndexFlatL2(dimension)
|
463 |
+
|
464 |
+
if torch.cuda.is_available():
|
465 |
+
# Use GPU for FAISS if available
|
466 |
+
res = faiss.StandardGpuResources()
|
467 |
+
self.index = faiss.index_cpu_to_gpu(res, 0, self.index)
|
468 |
+
|
469 |
+
self.index.add(np.array(embeddings).astype('float32'))
|
470 |
+
self.indexed_documents = documents
|
471 |
+
|
472 |
+
# Save index and documents
|
473 |
+
cpu_index = faiss.index_gpu_to_cpu(self.index) if torch.cuda.is_available() else self.index
|
474 |
+
faiss.write_index(cpu_index, str(self.index_path))
|
475 |
+
|
476 |
+
with open(self.documents_path, 'wb') as f:
|
477 |
+
pickle.dump(documents, f)
|
478 |
+
|
479 |
+
return True
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
480 |
|
481 |
def get_retriever(self, k: int = 10):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
482 |
if self.index is None:
|
483 |
if not self.load_index():
|
484 |
if not self.create_index():
|
485 |
+
raise ValueError("Unable to load or create index")
|
486 |
|
487 |
+
def retriever_function(query: str) -> list:
|
488 |
+
# Check cache first
|
489 |
+
cache_key = f"{query}_{k}"
|
490 |
+
if cache_key in self.response_cache:
|
491 |
+
return self.response_cache[cache_key]
|
492 |
|
493 |
+
query_embedding = self.encode(query)
|
494 |
+
|
|
|
|
|
|
|
495 |
distances, indices = self.index.search(
|
496 |
np.array([query_embedding]).astype('float32'),
|
497 |
k
|
498 |
)
|
499 |
+
|
500 |
+
results = [
|
501 |
+
self.indexed_documents[idx]
|
502 |
+
for idx in indices[0]
|
503 |
+
if idx != -1
|
504 |
+
]
|
505 |
+
|
506 |
+
# Cache the results
|
507 |
+
self.response_cache[cache_key] = results
|
508 |
return results
|
509 |
+
|
510 |
return retriever_function
|
511 |
|
512 |
+
# Initialize components
|
513 |
+
llm = ChatMistralAI(
|
514 |
+
model="mistral-large-latest",
|
515 |
+
mistral_api_key="QK0ZZpSxQbCEVgOLtI6FARQVmBYc6WGP",
|
516 |
+
temperature=0.1 # Lower temperature for faster responses
|
517 |
+
)
|
|
|
|
|
518 |
|
519 |
+
rag_loader = OptimizedRAGLoader()
|
520 |
+
retriever = rag_loader.get_retriever(k=5) # Reduced k for faster retrieval
|
521 |
|
522 |
+
# Cache for processed questions
|
523 |
+
question_cache = {}
|
|
|
|
|
524 |
|
525 |
def process_question(question: str) -> tuple[str, str]:
|
526 |
+
# Check cache first
|
527 |
+
if question in question_cache:
|
528 |
+
return question_cache[question]
|
529 |
+
|
530 |
relevant_docs = retriever(question)
|
531 |
context = "\n".join([doc.page_content for doc in relevant_docs])
|
532 |
+
|
533 |
prompt = prompt_template.format_messages(
|
534 |
context=context,
|
535 |
question=question
|
536 |
)
|
537 |
+
|
538 |
response = llm(prompt)
|
539 |
+
result = (response.content, context)
|
540 |
+
|
541 |
+
# Cache the result
|
542 |
+
question_cache[question] = result
|
543 |
+
return result
|
544 |
+
|
545 |
+
# Gradio interface with queue
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
546 |
with gr.Blocks(css=custom_css) as iface:
|
547 |
with gr.Column():
|
548 |
input_text = gr.Textbox(
|
|
|
551 |
lines=2,
|
552 |
elem_classes="rtl-text"
|
553 |
)
|
554 |
+
|
555 |
+
with gr.Row():
|
556 |
+
answer_box = gr.Textbox(
|
557 |
+
label="الإجابة",
|
558 |
+
lines=4,
|
559 |
+
elem_classes="rtl-text"
|
560 |
+
)
|
561 |
+
context_box = gr.Textbox(
|
562 |
+
label="السياق المستخدم",
|
563 |
+
lines=8,
|
564 |
+
elem_classes="rtl-text"
|
565 |
+
)
|
566 |
+
|
567 |
submit_btn = gr.Button("إرسال")
|
568 |
+
|
569 |
submit_btn.click(
|
570 |
+
fn=process_question,
|
571 |
inputs=input_text,
|
572 |
+
outputs=[answer_box, context_box],
|
573 |
+
api_name="predict"
|
574 |
)
|
575 |
|
576 |
+
# Launch with optimized settings
|
577 |
if __name__ == "__main__":
|
578 |
+
iface.queue(concurrency_count=3).launch(
|
579 |
+
share=True,
|
580 |
+
server_name="0.0.0.0",
|
581 |
+
server_port=7860,
|
582 |
+
enable_queue=True
|
583 |
+
)
|