File size: 22,401 Bytes
d9d7415
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
15c1ef1
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54c0e82
 
 
4739096
54c0e82
4739096
54c0e82
 
 
 
 
4739096
 
 
 
54c0e82
 
 
 
4739096
 
 
 
 
 
 
54c0e82
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
564c5be
4739096
 
564c5be
 
4739096
564c5be
 
 
 
 
 
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17917bf
4739096
17917bf
4739096
 
 
 
17917bf
4739096
 
 
 
 
 
17917bf
 
 
 
 
 
 
 
 
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e78ff28
564c5be
e78ff28
 
 
564c5be
 
 
 
 
 
 
e78ff28
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fca8e5d
 
 
4739096
 
 
 
fca8e5d
 
4739096
 
fca8e5d
4739096
 
 
fca8e5d
4739096
fca8e5d
4739096
 
 
 
fca8e5d
4739096
fca8e5d
4739096
 
fca8e5d
 
d9d7415
fca8e5d
d9d7415
 
 
 
 
 
 
 
 
 
fca8e5d
4739096
 
 
 
d9d7415
 
 
 
 
 
 
 
 
fca8e5d
 
4739096
 
54c0e82
 
 
4739096
d9d7415
54c0e82
 
d9d7415
 
 
 
 
 
54c0e82
 
 
 
 
 
 
 
 
 
 
 
 
d9d7415
 
 
 
 
 
54c0e82
4739096
 
54c0e82
d9d7415
 
 
fca8e5d
d9d7415
 
 
54c0e82
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d9d7415
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17917bf
 
 
 
 
4739096
 
 
 
17917bf
 
 
4739096
17917bf
 
 
 
 
 
 
 
 
 
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
17917bf
 
4739096
 
 
 
 
 
 
17917bf
 
4739096
 
 
 
 
 
fca8e5d
4739096
 
 
 
 
 
 
 
 
 
 
 
 
fca8e5d
4739096
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54c0e82
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
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
from llama_cpp import Llama
from typing import Optional, Dict, Any, List
import logging
import time
import os

from src.utils.config import RAGConfig
from src.router.query_router import QueryRouter
from src.prompts.dynamic_prompts import PromptManager

# ๋กœ๊น… ์„ค์ •
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)


class GGUFGenerator:
    """
    GGUF ๊ธฐ๋ฐ˜ Llama-3 ์ƒ์„ฑ๊ธฐ
    
    llama.cpp๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ GGUF ํฌ๋งท ๋ชจ๋ธ์„ ๋กœ๋“œํ•˜๊ณ 
    ์ž…์ฐฐ ๊ด€๋ จ ์งˆ์˜์‘๋‹ต์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
    """
    
    def __init__(
        self,
        model_path: str,
        n_gpu_layers: int = 0,
        n_ctx: int = 8192,
        n_threads: int = 8,
        config = None,
        max_new_tokens: int = 256,
        temperature: float = 0.7,
        top_p: float = 0.9,
        system_prompt: str = "๋‹น์‹ ์€ RFP(์ œ์•ˆ์š”์ฒญ์„œ) ๋ถ„์„ ๋ฐ ์š”์•ฝ ์ „๋ฌธ๊ฐ€์ž…๋‹ˆ๋‹ค."
    ):
        """
        ์ƒ์„ฑ๊ธฐ ์ดˆ๊ธฐํ™”
        
        Args:
            model_path: GGUF ๋ชจ๋ธ ํŒŒ์ผ ๊ฒฝ๋กœ
            n_gpu_layers: GPU์— ์˜ฌ๋ฆด ๋ ˆ์ด์–ด ์ˆ˜ (0 = CPU๋งŒ, 35 = ์ „์ฒด GPU)
            n_ctx: ์ตœ๋Œ€ ์ปจํ…์ŠคํŠธ ๊ธธ์ด
            n_threads: CPU ์Šค๋ ˆ๋“œ ์ˆ˜
            max_new_tokens: ์ตœ๋Œ€ ์ƒ์„ฑ ํ† ํฐ ์ˆ˜
            temperature: ์ƒ์„ฑ ๋‹ค์–‘์„ฑ (0.0~1.0)
            top_p: Nucleus sampling ํŒŒ๋ผ๋ฏธํ„ฐ
            system_prompt: ์‹œ์Šคํ…œ ํ”„๋กฌํ”„ํŠธ
        """
        self.config = config or RAGConfig() 
        self.model_path = model_path
        self.n_gpu_layers = n_gpu_layers
        self.n_ctx = n_ctx
        self.n_threads = n_threads
        self.max_new_tokens = max_new_tokens
        self.temperature = temperature
        self.top_p = top_p
        self.system_prompt = system_prompt
        
        # ๋ชจ๋ธ (๋‚˜์ค‘์— ๋กœ๋“œ)
        self.model = None
        
        logger.info(f"GGUFGenerator ์ดˆ๊ธฐํ™” ์™„๋ฃŒ")
    
    def load_model(self) -> None:
        """
        GGUF ๋ชจ๋ธ ๋กœ๋“œ
        
        ๋กœ์ง:
        1. USE_MODEL_HUB ํ™•์ธ
        2-A. True โ†’ Hugging Face Hub์—์„œ ๋‹ค์šด๋กœ๋“œ
        2-B. False โ†’ ๋กœ์ปฌ ํŒŒ์ผ ์‚ฌ์šฉ
        3. ๋ชจ๋ธ ๋กœ๋“œ
        """
        
        # ์ค‘๋ณต ๋กœ๋“œ ๋ฐฉ์ง€
        if self.model is not None:
            logger.info("๋ชจ๋ธ์ด ์ด๋ฏธ ๋กœ๋“œ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.")
            return
        
        try:
            # Config์—์„œ USE_MODEL_HUB ํ™•์ธ (์—†์œผ๋ฉด True ๊ธฐ๋ณธ๊ฐ’)
            use_model_hub = getattr(self.config, 'USE_MODEL_HUB', True)
            
            # Model Hub ์‚ฌ์šฉ ์—ฌ๋ถ€์— ๋”ฐ๋ผ ๊ฒฝ๋กœ ๊ฒฐ์ •
            if use_model_hub:
                # === Model Hub์—์„œ ๋‹ค์šด๋กœ๋“œ ===
                model_hub_repo = getattr(self.config, 'MODEL_HUB_REPO', 'beomi/Llama-3-Open-Ko-8B-gguf')
                model_hub_filename = getattr(self.config, 'MODEL_HUB_FILENAME', 'ggml-model-Q4_K_M.gguf')
                model_cache_dir = getattr(self.config, 'MODEL_CACHE_DIR', '.cache/models')
                
                logger.info(f"๐Ÿ“ฅ Model Hub์—์„œ ๋‹ค์šด๋กœ๋“œ: {model_hub_repo}")
                
                from huggingface_hub import hf_hub_download
                
                model_path = hf_hub_download(
                    repo_id=model_hub_repo,
                    filename=model_hub_filename,
                    cache_dir=model_cache_dir,
                    local_dir=model_cache_dir,
                    local_dir_use_symlinks=False  # ์‹ฌ๋ณผ๋ฆญ ๋งํฌ ๋Œ€์‹  ์‹ค์ œ ๋ณต์‚ฌ
                )
                
                logger.info(f"โœ… ๋‹ค์šด๋กœ๋“œ ์™„๋ฃŒ: {model_path}")
                
            else:
                # === ๋กœ์ปฌ ํŒŒ์ผ ์‚ฌ์šฉ ===
                model_path = self.model_path  # ์ƒ์„ฑ์ž์—์„œ ๋ฐ›์€ ๊ฒฝ๋กœ ์‚ฌ์šฉ
                
                if not os.path.exists(model_path):
                    raise FileNotFoundError(
                        f"โŒ ๋กœ์ปฌ ๋ชจ๋ธ ํŒŒ์ผ์„ ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค: {model_path}\n"
                        f"   USE_MODEL_HUB=true๋กœ ์„ค์ •ํ•˜๊ฑฐ๋‚˜ ๋ชจ๋ธ ํŒŒ์ผ์„ ์ค€๋น„ํ•˜์„ธ์š”."
                    )
                
                logger.info(f"๐Ÿ“‚ ๋กœ์ปฌ ๋ชจ๋ธ ์‚ฌ์šฉ: {model_path}")
            
            # === ๊ณตํ†ต: ๋ชจ๋ธ ๋กœ๋“œ ===
            logger.info(f"๐Ÿš€ GGUF ๋ชจ๋ธ ๋กœ๋“œ ์ค‘...")
            logger.info(f"   GPU ๋ ˆ์ด์–ด: {self.n_gpu_layers}")
            logger.info(f"   ์ปจํ…์ŠคํŠธ: {self.n_ctx}")
            
            self.model = Llama(
                model_path=model_path,
                n_gpu_layers=self.n_gpu_layers,
                n_ctx=self.n_ctx,
                n_threads=self.n_threads,
                verbose=True,  # โœ… ๋””๋ฒ„๊ทธ ๋กœ๊ทธ ํ™œ์„ฑํ™”
            )
            
            # โœ… ์‹ค์ œ ์ ์šฉ๋œ n_ctx ํ™•์ธ
            actual_n_ctx = self.model.n_ctx()
            logger.info("โœ… GGUF ๋ชจ๋ธ ๋กœ๋“œ ์™„๋ฃŒ!")
            logger.info(f"   - ์„ค์ •ํ•œ n_ctx: {self.n_ctx}")
            logger.info(f"   - ์‹ค์ œ n_ctx: {actual_n_ctx}")
            
            if actual_n_ctx < self.n_ctx:
                logger.warning(f"โš ๏ธ n_ctx๊ฐ€ ์˜ˆ์ƒ๋ณด๋‹ค ์ž‘์Šต๋‹ˆ๋‹ค: {actual_n_ctx} < {self.n_ctx}")
                logger.warning(f"   ๋ฉ”๋ชจ๋ฆฌ ๋ถ€์กฑ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. n_gpu_layers๋ฅผ ์ค„์—ฌ๋ณด์„ธ์š”.")
            
        except FileNotFoundError as e:
            logger.error(f"โŒ ๋ชจ๋ธ ํŒŒ์ผ์„ ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค: {e}")
            raise
        except Exception as e:
            logger.error(f"โŒ ๋ชจ๋ธ ๋กœ๋“œ ์‹คํŒจ: {e}")
            raise RuntimeError(f"๋ชจ๋ธ ๋กœ๋“œ ์ค‘ ์˜ค๋ฅ˜ ๋ฐœ์ƒ: {e}")
    
    def format_prompt(
        self,
        question: str,
        context: Optional[str] = None,
        system_prompt: Optional[str] = None
    ) -> str:
        """
        GGUF ๋ชจ๋ธ์šฉ ๊ฐ„๋‹จํ•œ ํ”„๋กฌํ”„ํŠธ ํฌ๋งทํŒ…
        
        Llama-3 ํŠน์ˆ˜ ํ† ํฐ ๋Œ€์‹  ์ˆœ์ˆ˜ ํ…์ŠคํŠธ ๊ธฐ๋ฐ˜ ํ…œํ”Œ๋ฆฟ ์‚ฌ์šฉ
        """
        # ์‹œ์Šคํ…œ ํ”„๋กฌํ”„ํŠธ ์„ค์ •
        if system_prompt is None:
            system_prompt = self.system_prompt
        
        # ์ปจํ…์ŠคํŠธ ํฌํ•จ ์—ฌ๋ถ€
        if context is not None:
            user_message = f"์ฐธ๊ณ  ๋ฌธ์„œ:\n{context}\n\n์งˆ๋ฌธ: {question}"
        else:
            user_message = question
        
        # ๊ฐ„๋‹จํ•œ ํ•œ๊ตญ์–ด ํ…œํ”Œ๋ฆฟ (ํŠน์ˆ˜ ํ† ํฐ ์—†์Œ)
        formatted_prompt = f"""### ์‹œ์Šคํ…œ
{system_prompt}

### ์‚ฌ์šฉ์ž
{user_message}

### ๋‹ต๋ณ€
"""
        
        return formatted_prompt
    
    def generate(
        self,
        prompt: str,
        max_new_tokens: Optional[int] = None,
        temperature: Optional[float] = None,
        top_p: Optional[float] = None,
    ) -> str:
        """
        ํ”„๋กฌํ”„ํŠธ๋ฅผ ์ž…๋ ฅ๋ฐ›์•„ ์‘๋‹ต ์ƒ์„ฑ
        
        Args:
            prompt: ํฌ๋งท๋œ ํ”„๋กฌํ”„ํŠธ
            max_new_tokens: ์ตœ๋Œ€ ์ƒ์„ฑ ํ† ํฐ ์ˆ˜
            temperature: ์ƒ์„ฑ ๋‹ค์–‘์„ฑ
            top_p: Nucleus sampling
        
        Returns:
            ์ƒ์„ฑ๋œ ์‘๋‹ต ํ…์ŠคํŠธ
        
        Raises:
            RuntimeError: ๋ชจ๋ธ์ด ๋กœ๋“œ๋˜์ง€ ์•Š์€ ๊ฒฝ์šฐ
        """
        # ๋ชจ๋ธ ๋กœ๋“œ ํ™•์ธ
        if self.model is None:
            raise RuntimeError(
                "๋ชจ๋ธ์ด ๋กœ๋“œ๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. load_model()์„ ๋จผ์ € ํ˜ธ์ถœํ•˜์„ธ์š”."
            )
        
        # ํŒŒ๋ผ๋ฏธํ„ฐ ์„ค์ •
        if max_new_tokens is None:
            max_new_tokens = self.max_new_tokens
        if temperature is None:
            temperature = self.temperature
        if top_p is None:
            top_p = self.top_p
        
        try:
            logger.info(f"๐Ÿ”„ ์ƒ์„ฑ ์‹œ์ž‘ (max_tokens={max_new_tokens}, temp={temperature})")
            start_time = time.time()
            
            # ์ƒ์„ฑ
            output = self.model(
                prompt,
                max_tokens=max_new_tokens,
                temperature=temperature,
                top_p=top_p,
                echo=False,  # ํ”„๋กฌํ”„ํŠธ ๋ฐ˜๋ณต ์•ˆ ํ•จ
                stop=[
                    # ๊ตฌ๋ถ„์ž
                    "###", "\n\n###", 
                    "### ์‚ฌ์šฉ์ž", "\n์‚ฌ์šฉ์ž:", 
                    "</s>",
                    # ๋ฉ”ํƒ€ ํ…์ŠคํŠธ ์ฐจ๋‹จ
                    "ํ•œ๊ตญ์–ด ๋‹ต๋ณ€", "ํ•œ๊ตญ์–ด๋กœ ๋‹ต๋ณ€", "์ง€์นจ:",
                    "๋ฌธ์žฅ", "(๋ฌธ์žฅ",
                    # โœ… ์งˆ๋ฌธ ํŒจํ„ด ์ฐจ๋‹จ (๋‹ต๋ณ€ ํ›„ ์งˆ๋ฌธ ์ƒ์„ฑ ๋ฐฉ์ง€)
                    "\n\n",  # ๋‹จ๋ฝ ๊ตฌ๋ถ„
                    "?",     # ์งˆ๋ฌธ ๊ธฐํ˜ธ
                    "์š”?", "๊นŒ?", "๋‚˜์š”?", "์Šต๋‹ˆ๊นŒ?"  # ์งˆ๋ฌธ ์–ด๋ฏธ
                ],
            )
            
            elapsed = time.time() - start_time
            logger.info(f"โœ… ์ƒ์„ฑ ์™„๋ฃŒ: {elapsed:.2f}์ดˆ")
            
            # ์‘๋‹ต ์ถ”์ถœ
            response = output['choices'][0]['text'].strip()
            
            logger.info(f"๐Ÿ“ ์‘๋‹ต ๊ธธ์ด: {len(response)} ๊ธ€์ž")
            return response
            
        except Exception as e:
            logger.error(f"โŒ ์ƒ์„ฑ ์ค‘ ์˜ค๋ฅ˜ ๋ฐœ์ƒ: {e}")
            raise RuntimeError(f"ํ…์ŠคํŠธ ์ƒ์„ฑ ์‹คํŒจ: {e}")
    
    def chat(
        self,
        question: str,
        context: Optional[str] = None,
        system_prompt=None,
        **kwargs
    ) -> str:
        """
        ์งˆ๋ฌธ์— ๋Œ€ํ•œ ์‘๋‹ต ์ƒ์„ฑ (ํ†ตํ•ฉ ๋ฉ”์„œ๋“œ)
        
        Args:
            question: ์‚ฌ์šฉ์ž ์งˆ๋ฌธ
            context: ์„ ํƒ์  ์ปจํ…์ŠคํŠธ
            system_prompt: ์„ ํƒ์  ์‹œ์Šคํ…œ ํ”„๋กฌํ”„ํŠธ
            **kwargs: generate() ๋ฉ”์„œ๋“œ์— ์ „๋‹ฌ๋  ์ถ”๊ฐ€ ํŒŒ๋ผ๋ฏธํ„ฐ
        
        Returns:
            ์ƒ์„ฑ๋œ ์‘๋‹ต
        """
        # ํ”„๋กฌํ”„ํŠธ ํฌ๋งทํŒ…
        prompt = self.format_prompt(
            question=question,
            context=context,
            system_prompt=system_prompt
        )
        
        # ์‘๋‹ต ์ƒ์„ฑ
        response = self.generate(prompt, **kwargs)
        
        return response


class GGUFRAGPipeline:
    """
    GGUF ์ƒ์„ฑ๊ธฐ + RAG ํ†ตํ•ฉ ํŒŒ์ดํ”„๋ผ์ธ
    
    chatbot_app.py์™€ ํ˜ธํ™˜๋˜๋Š” ์ธํ„ฐํŽ˜์ด์Šค ์ œ๊ณต
    """
    
    def __init__(
        self,
        config=None,
        model: str = None,  # ํ˜ธํ™˜์„ฑ์šฉ (์‚ฌ์šฉ ์•ˆ ํ•จ)
        top_k: int = None,
        # GPU ์„ค์ • (์„ ํƒ์ , config ์˜ค๋ฒ„๋ผ์ด๋“œ)
        n_gpu_layers: int = None,
        n_ctx: int = None,
        n_threads: int = None,
        max_new_tokens: int = None,
        temperature: float = None,
        top_p: float = None,
        search_mode: str = None,
        alpha: float = None
    ):
        """
        ์ดˆ๊ธฐํ™”
        
        Args:
            config: RAGConfig ๊ฐ์ฒด
            model: ๋ชจ๋ธ ์ด๋ฆ„ (์‚ฌ์šฉ ์•ˆ ํ•จ, ํ˜ธํ™˜์„ฑ์šฉ)
            top_k: ๊ธฐ๋ณธ ๊ฒ€์ƒ‰ ๋ฌธ์„œ ์ˆ˜
            n_gpu_layers: GPU ๋ ˆ์ด์–ด ์ˆ˜ (config ์˜ค๋ฒ„๋ผ์ด๋“œ)
            n_ctx: ์ปจํ…์ŠคํŠธ ๊ธธ์ด (config ์˜ค๋ฒ„๋ผ์ด๋“œ)
            n_threads: CPU ์Šค๋ ˆ๋“œ ์ˆ˜ (config ์˜ค๋ฒ„๋ผ์ด๋“œ)
            max_new_tokens: ์ตœ๋Œ€ ์ƒ์„ฑ ํ† ํฐ (config ์˜ค๋ฒ„๋ผ์ด๋“œ)
            temperature: ์ƒ์„ฑ ๋‹ค์–‘์„ฑ (config ์˜ค๋ฒ„๋ผ์ด๋“œ)
            top_p: Nucleus sampling (config ์˜ค๋ฒ„๋ผ์ด๋“œ)
            search_mode: ๊ฒ€์ƒ‰ ๋ชจ๋“œ
            alpha: ์ž„๋ฒ ๋”ฉ ๊ฐ€์ค‘์น˜
        """
        self.config = config or RAGConfig()
        
        # Config์—์„œ ๊ธฐ๋ณธ๊ฐ’ ๊ฐ€์ ธ์˜ค๊ธฐ (์—†์œผ๋ฉด fallback)
        self.top_k = top_k or getattr(self.config, 'DEFAULT_TOP_K', 10)
        
        # ๊ฒ€์ƒ‰ ์„ค์ •
        self.search_mode = search_mode or getattr(self.config, 'DEFAULT_SEARCH_MODE', 'hybrid_rerank')
        self.alpha = alpha if alpha is not None else getattr(self.config, 'DEFAULT_ALPHA', 0.5)
        
        # Retriever ์ดˆ๊ธฐํ™” (RAGRetriever ์‚ฌ์šฉ)
        logger.info("RAGRetriever ์ดˆ๊ธฐํ™” ์ค‘...")
        from src.retriever.retriever import RAGRetriever
        self.retriever = RAGRetriever(config=self.config)
        
        # GGUF ์„ค์ • (ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์ฃผ์–ด์ง€๋ฉด config ์˜ค๋ฒ„๋ผ์ด๋“œ, ์—†์œผ๋ฉด ๊ธฐ๋ณธ๊ฐ’)
        gguf_n_gpu_layers = n_gpu_layers if n_gpu_layers is not None else getattr(self.config, 'GGUF_N_GPU_LAYERS', 35)
        gguf_n_ctx = n_ctx if n_ctx is not None else getattr(self.config, 'GGUF_N_CTX', 2048)
        gguf_n_threads = n_threads if n_threads is not None else getattr(self.config, 'GGUF_N_THREADS', 4)
        gguf_max_new_tokens = max_new_tokens if max_new_tokens is not None else getattr(self.config, 'GGUF_MAX_NEW_TOKENS', 512)
        gguf_temperature = temperature if temperature is not None else getattr(self.config, 'GGUF_TEMPERATURE', 0.7)
        gguf_top_p = top_p if top_p is not None else getattr(self.config, 'GGUF_TOP_P', 0.9)
        
        # ๋ชจ๋ธ ๊ฒฝ๋กœ (fallback)
        gguf_model_path = getattr(self.config, 'GGUF_MODEL_PATH', '.cache/models/llama-3-ko-8b.gguf')
        
        # ์‹œ์Šคํ…œ ํ”„๋กฌํ”„ํŠธ (fallback)
        system_prompt = getattr(self.config, 'SYSTEM_PROMPT', '๋‹น์‹ ์€ ํ•œ๊ตญ ๊ณต๊ณต๊ธฐ๊ด€ ์‚ฌ์—…์ œ์•ˆ์„œ ๋ถ„์„ ์ „๋ฌธ๊ฐ€์ž…๋‹ˆ๋‹ค.')
        
        # GGUFGenerator ์ดˆ๊ธฐํ™”
        logger.info("GGUFGenerator ์ดˆ๊ธฐํ™” ์ค‘...")
        logger.info(f"   GPU ๋ ˆ์ด์–ด: {gguf_n_gpu_layers}")
        logger.info(f"   ์ปจํ…์ŠคํŠธ: {gguf_n_ctx}")
        logger.info(f"   ์Šค๋ ˆ๋“œ: {gguf_n_threads}")
        logger.info(f"   ๋ชจ๋ธ ๊ฒฝ๋กœ: {gguf_model_path}")
        
        self.generator = GGUFGenerator(
            model_path=gguf_model_path,
            n_gpu_layers=gguf_n_gpu_layers,
            n_ctx=gguf_n_ctx,
            n_threads=gguf_n_threads,
            config=self.config,
            max_new_tokens=gguf_max_new_tokens,
            temperature=gguf_temperature,
            top_p=gguf_top_p,
            system_prompt=system_prompt
        )
        
        # ๋ชจ๋ธ ๋กœ๋“œ (์‹œ๊ฐ„ ์†Œ์š”)
        logger.info("GGUF ๋ชจ๋ธ ๋กœ๋“œ ์ค‘...")
        self.generator.load_model()
        
        # ๋Œ€ํ™” ํžˆ์Šคํ† ๋ฆฌ
        self.chat_history: List[Dict] = []
        
        # ๋งˆ์ง€๋ง‰ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ ์ €์žฅ (sources ๋ฐ˜ํ™˜์šฉ)
        self._last_retrieved_docs = []
        
        logger.info("โœ… GGUFRAGPipeline ์ดˆ๊ธฐํ™” ์™„๋ฃŒ")
        logger.info(f"   - ๊ฒ€์ƒ‰ ๋ชจ๋“œ: {self.search_mode}")
        logger.info(f"   - ๊ธฐ๋ณธ top_k: {self.top_k}")
    
    def _retrieve_and_format(self, query: str) -> str:
        """๊ฒ€์ƒ‰ ์ˆ˜ํ–‰ ๋ฐ ์ปจํ…์ŠคํŠธ ํฌ๋งทํŒ…"""
        # ๊ฒ€์ƒ‰ ๋ชจ๋“œ์— ๋”ฐ๋ผ ๋ฌธ์„œ ๊ฒ€์ƒ‰ (RAGRetriever ๋ฉ”์„œ๋“œ ์‚ฌ์šฉ)
        if self.search_mode == "embedding":
            docs = self.retriever.search(query, top_k=self.top_k)
        elif self.search_mode == "embedding_rerank":
            docs = self.retriever.search_with_rerank(query, top_k=self.top_k)
        elif self.search_mode == "hybrid":
            docs = self.retriever.hybrid_search(
                query, top_k=self.top_k, alpha=self.alpha
            )
        elif self.search_mode == "hybrid_rerank":
            docs = self.retriever.hybrid_search_with_rerank(
                query, top_k=self.top_k, alpha=self.alpha
            )
        else:
            docs = self.retriever.search(query, top_k=self.top_k)
        
        # ๋งˆ์ง€๋ง‰ ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ ์ €์žฅ
        self._last_retrieved_docs = docs
        
        # ์ปจํ…์ŠคํŠธ ํฌ๋งทํŒ…
        return self._format_context(docs)
    
    def _format_context(self, retrieved_docs: list) -> str:
        """
        ๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ๋ฅผ ์ปจํ…์ŠคํŠธ๋กœ ๋ณ€ํ™˜
        
        ์ปจํ…์ŠคํŠธ๊ฐ€ ๋„ˆ๋ฌด ๊ธธ๋ฉด ์ž๋™์œผ๋กœ ์ค„์ž„ (ํ† ํฐ ์ œํ•œ ๋Œ€์‘)
        """
        if not retrieved_docs:
            return "๊ด€๋ จ ๋ฌธ์„œ๋ฅผ ์ฐพ์„ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค."
        
        context_parts = []
        max_context_chars = 8000  # ๋Œ€๋žต 2000 ํ† ํฐ ์ •๋„ (์—ฌ์œ  ์žˆ๊ฒŒ)
        
        current_length = 0
        for i, doc in enumerate(retrieved_docs, 1):
            doc_text = f"[๋ฌธ์„œ {i}]\n{doc['content']}\n"
            doc_length = len(doc_text)
            
            # ์ปจํ…์ŠคํŠธ ๊ธธ์ด ์ฒดํฌ
            if current_length + doc_length > max_context_chars:
                logger.warning(f"โš ๏ธ ์ปจํ…์ŠคํŠธ ๊ธธ์ด ์ œํ•œ: {i-1}๊ฐœ ๋ฌธ์„œ๋งŒ ์‚ฌ์šฉ (์ตœ๋Œ€ {max_context_chars}์ž)")
                break
            
            context_parts.append(doc_text)
            current_length += doc_length
        
        return "\n".join(context_parts)
    
    def _format_sources(self, retrieved_docs: list) -> list:
        """๊ฒ€์ƒ‰๋œ ๋ฌธ์„œ๋ฅผ sources ํ˜•์‹์œผ๋กœ ๋ณ€ํ™˜"""
        sources = []
        for doc in retrieved_docs:
            source_info = {
                'content': doc['content'],
                'metadata': doc['metadata'],
                'filename': doc.get('filename', 'N/A'),
                'organization': doc.get('organization', 'N/A')
            }
            
            # ๊ฒ€์ƒ‰ ๋ชจ๋“œ์— ๋”ฐ๋ผ ์ ์ˆ˜ ํ•„๋“œ๊ฐ€ ๋‹ค๋ฆ„
            if 'rerank_score' in doc:
                source_info['score'] = doc['rerank_score']
                source_info['score_type'] = 'rerank'
            elif 'hybrid_score' in doc:
                source_info['score'] = doc['hybrid_score']
                source_info['score_type'] = 'hybrid'
            elif 'relevance_score' in doc:
                source_info['score'] = doc['relevance_score']
                source_info['score_type'] = 'embedding'
            else:
                source_info['score'] = 0
                source_info['score_type'] = 'unknown'
            
            sources.append(source_info)
        
        return sources
    
    def _estimate_usage(self, query: str, answer: str) -> dict:
        """ํ† ํฐ ์‚ฌ์šฉ๋Ÿ‰ ์ถ”์ •"""
        # ๊ฐ„๋‹จํ•œ ๋‹จ์–ด ์ˆ˜ ๊ธฐ๋ฐ˜ ์ถ”์ •
        prompt_tokens = len(query.split()) * 2
        completion_tokens = len(answer.split()) * 2
        
        return {
            'total_tokens': prompt_tokens + completion_tokens,
            'prompt_tokens': prompt_tokens,
            'completion_tokens': completion_tokens
        }
    
    def generate_answer(
        self,
        query: str,
        top_k: int = None,
        search_mode: str = None,
        alpha: float = None
    ) -> dict:
        """
        ๋‹ต๋ณ€ ์ƒ์„ฑ (chatbot_app.py ํ˜ธํ™˜ ๋ฉ”์ธ ๋ฉ”์„œ๋“œ)
        
        Args:
            query: ์งˆ๋ฌธ
            top_k: ๊ฒ€์ƒ‰ํ•  ๋ฌธ์„œ ์ˆ˜
            search_mode: ๊ฒ€์ƒ‰ ๋ชจ๋“œ
            alpha: ์ž„๋ฒ ๋”ฉ ๊ฐ€์ค‘์น˜
        
        Returns:
            dict: answer, sources, search_mode, usage, elapsed_time, used_retrieval
        """
        try:
            start_time = time.time()
            
            # ํŒŒ๋ผ๋ฏธํ„ฐ ์„ค์ • (๊ฒ€์ƒ‰ ์ „์— ๋จผ์ € ์„ค์ •)
            if top_k is not None:
                self.top_k = top_k
            if search_mode is not None:
                self.search_mode = search_mode
            if alpha is not None:
                self.alpha = alpha

            # ===== Router๋กœ ๊ฒ€์ƒ‰ ์—ฌ๋ถ€ ๊ฒฐ์ • =====
            router = QueryRouter()
            classification = router.classify(query)
            query_type = classification['type']  # 'greeting'/'thanks'/'document'/'out_of_scope'
            
            logger.info(f"๐Ÿ“ ๋ถ„๋ฅ˜: {query_type} "
                f"(์‹ ๋ขฐ๋„: {classification['confidence']:.2f})")
            
            # 2. ํƒ€์ž…๋ณ„ ์ฒ˜๋ฆฌ
            if query_type in ['greeting', 'thanks', 'out_of_scope']:
                # ๊ฒ€์ƒ‰ ์Šคํ‚ต
                context = None
                used_retrieval = False
                self._last_retrieved_docs = []
                
                # ๋™์  ํ”„๋กฌํ”„ํŠธ ์„ ํƒ (GGUF์šฉ)
                system_prompt = PromptManager.get_prompt(query_type, model_type="gguf")
                logger.info(f"โญ๏ธ RAG ์Šคํ‚ต: {query_type}")
            
            elif query_type == 'document':
                # RAG ์ˆ˜ํ–‰
                context = self._retrieve_and_format(query)
                used_retrieval = True
                
                # ๋™์  ํ”„๋กฌํ”„ํŠธ (GGUF์šฉ, context ํฌํ•จ)
                system_prompt = PromptManager.get_prompt('document', model_type="gguf")
                logger.info(f"๐Ÿ” RAG ์ˆ˜ํ–‰: {len(self._last_retrieved_docs)}๊ฐœ ๋ฌธ์„œ")
            
            # 3. ๋‹ต๋ณ€ ์ƒ์„ฑ (system_prompt ์ „๋‹ฌ)
            answer = self.generator.chat(
                question=query,
                context=context,
                system_prompt=system_prompt
            )
            
            elapsed_time = time.time() - start_time
            
            # ๋Œ€ํ™” ํžˆ์Šคํ† ๋ฆฌ์— ์ถ”๊ฐ€
            self.chat_history.append({"role": "user", "content": query})
            self.chat_history.append({"role": "assistant", "content": answer})
            
            # ๊ฒฐ๊ณผ ๋ฐ˜ํ™˜ (RAGPipeline๊ณผ ๋™์ผ ํ˜•์‹)
            return {
                'answer': answer,
                'sources': self._format_sources(self._last_retrieved_docs),
                'used_retrieval': used_retrieval,
                'query_type': query_type,
                'search_mode': self.search_mode if used_retrieval else 'direct',
                'routing_info': classification,
                'elapsed_time': elapsed_time,
                'usage': self._estimate_usage(query, answer)
            }
        
        except Exception as e:
            logger.error(f"โŒ ๋‹ต๋ณ€ ์ƒ์„ฑ ์‹คํŒจ: {e}")
            import traceback
            traceback.print_exc()
            raise RuntimeError(f"๋‹ต๋ณ€ ์ƒ์„ฑ ์‹คํŒจ: {str(e)}") from e
    
    def chat(self, query: str) -> str:
        """๊ฐ„๋‹จํ•œ ๋Œ€ํ™” ์ธํ„ฐํŽ˜์ด์Šค"""
        result = self.generate_answer(query)
        return result['answer']
    
    def clear_history(self):
        """๋Œ€ํ™” ํžˆ์Šคํ† ๋ฆฌ ์ดˆ๊ธฐํ™”"""
        self.chat_history = []
        logger.info("๐Ÿ—‘๏ธ ๋Œ€ํ™” ํžˆ์Šคํ† ๋ฆฌ๊ฐ€ ์ดˆ๊ธฐํ™”๋˜์—ˆ์Šต๋‹ˆ๋‹ค.")
    
    def get_history(self) -> List[Dict]:
        """๋Œ€ํ™” ํžˆ์Šคํ† ๋ฆฌ ๋ฐ˜ํ™˜"""
        return self.chat_history.copy()
    
    def set_search_config(
        self,
        search_mode: str = None,
        top_k: int = None,
        alpha: float = None
    ):
        """๊ฒ€์ƒ‰ ์„ค์ • ๋ณ€๊ฒฝ"""
        if search_mode is not None:
            self.search_mode = search_mode
        if top_k is not None:
            self.top_k = top_k
        if alpha is not None:
            self.alpha = alpha
        
        logger.info(
            f"๐Ÿ”ง ๊ฒ€์ƒ‰ ์„ค์ • ๋ณ€๊ฒฝ: mode={self.search_mode}, "
            f"top_k={self.top_k}, alpha={self.alpha}"
        )