File size: 12,275 Bytes
9f620cb
 
 
 
 
 
 
 
 
 
 
3cf7a11
 
 
 
 
 
 
9f620cb
3cf7a11
9f620cb
 
 
3cf7a11
 
9f620cb
 
 
3cf7a11
 
 
 
 
 
9f620cb
 
 
3cf7a11
 
9f620cb
 
3cf7a11
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9f620cb
3cf7a11
9f620cb
 
3cf7a11
 
 
 
9f620cb
3cf7a11
 
 
 
 
 
9f620cb
 
3cf7a11
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9f620cb
 
3dad413
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3cf7a11
 
 
 
 
9f620cb
3cf7a11
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9f620cb
089275b
 
 
 
f5717d7
9f620cb
 
 
3cf7a11
9f620cb
3cf7a11
 
 
 
 
9f620cb
3cf7a11
 
9f620cb
 
 
 
3cf7a11
 
 
 
 
 
 
 
 
9f620cb
3cf7a11
9f620cb
 
3cf7a11
a70531f
3cf7a11
 
caa8ad3
3cf7a11
 
9f620cb
3cf7a11
77e13a3
9f620cb
3cf7a11
 
9f620cb
52b50a9
 
 
 
 
 
 
 
 
 
 
0022f8a
 
 
 
3cf7a11
0022f8a
 
3cf7a11
0022f8a
 
 
 
3cf7a11
0022f8a
 
3cf7a11
0022f8a
 
 
 
df0e374
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0022f8a
 
 
 
 
 
df0e374
0022f8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
df0e374
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3cf7a11
df0e374
 
 
 
 
 
 
 
 
 
 
3cf7a11
df0e374
3cf7a11
df0e374
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3cf7a11
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
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
import threading
from queue import Queue
import concurrent.futures

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)
        
        # 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 = 5):
        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=0.1  # Lower temperature for faster responses
)

rag_loader = OptimizedRAGLoader()
retriever = rag_loader.get_retriever(k=10)  # 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، قم بتوضيح ذلك.
    أجب بشكل موجز ودقيق."""),
    ("human", "{question}")
])

# def process_question(question: str) -> tuple[str, str]:
#     # Check cache first
#     if question in question_cache:
#         return 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
#     )
    
#     response = llm(prompt)
#     result = (response.content, context)
    
#     # Cache the result
#     question_cache[question] = result
#     return result

# # 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"
#         )

# 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 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  
    )