File size: 39,895 Bytes
519b145
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
"""
LLM integration for natural answer generation.
Supports OpenAI GPT, Anthropic Claude, Ollama, Hugging Face Inference API, Local Hugging Face models, and API mode.
"""
import os
import re
import json
import sys
import traceback
import logging
import time
from typing import List, Dict, Any, Optional
try:
    from dotenv import load_dotenv
    load_dotenv()
except ImportError:
    pass  # dotenv is optional

logger = logging.getLogger(__name__)

# Import download progress tracker (optional)
try:
    from .download_progress import get_progress_tracker, DownloadProgress
    PROGRESS_TRACKER_AVAILABLE = True
except ImportError:
    PROGRESS_TRACKER_AVAILABLE = False
    logger.warning("Download progress tracker not available")

# LLM Provider types
LLM_PROVIDER_OPENAI = "openai"
LLM_PROVIDER_ANTHROPIC = "anthropic"
LLM_PROVIDER_OLLAMA = "ollama"
LLM_PROVIDER_HUGGINGFACE = "huggingface"  # Hugging Face Inference API
LLM_PROVIDER_LOCAL = "local"  # Local Hugging Face Transformers model
LLM_PROVIDER_API = "api"  # API mode - call HF Spaces API
LLM_PROVIDER_NONE = "none"

# Get provider from environment (default to local Qwen if none provided)
DEFAULT_LLM_PROVIDER = os.environ.get("DEFAULT_LLM_PROVIDER", LLM_PROVIDER_LOCAL).lower()
env_provider = os.environ.get("LLM_PROVIDER", "").strip().lower()
LLM_PROVIDER = env_provider or DEFAULT_LLM_PROVIDER


class LLMGenerator:
    """Generate natural language answers using LLMs."""
    
    def __init__(self, provider: Optional[str] = None):
        """
        Initialize LLM generator.
        
        Args:
            provider: LLM provider ('openai', 'anthropic', 'ollama', 'local', 'huggingface', 'api', or None for auto-detect).
        """
        self.provider = provider or LLM_PROVIDER
        self.client = None
        self.local_model = None
        self.local_tokenizer = None
        self.api_base_url = None
        self._initialize_client()
    
    def _initialize_client(self):
        """Initialize LLM client based on provider."""
        if self.provider == LLM_PROVIDER_OPENAI:
            try:
                import openai
                api_key = os.environ.get("OPENAI_API_KEY")
                if api_key:
                    self.client = openai.OpenAI(api_key=api_key)
                    print("✅ OpenAI client initialized")
                else:
                    print("⚠️ OPENAI_API_KEY not found, OpenAI disabled")
            except ImportError:
                print("⚠️ openai package not installed, install with: pip install openai")
        
        elif self.provider == LLM_PROVIDER_ANTHROPIC:
            try:
                import anthropic
                api_key = os.environ.get("ANTHROPIC_API_KEY")
                if api_key:
                    self.client = anthropic.Anthropic(api_key=api_key)
                    print("✅ Anthropic client initialized")
                else:
                    print("⚠️ ANTHROPIC_API_KEY not found, Anthropic disabled")
            except ImportError:
                print("⚠️ anthropic package not installed, install with: pip install anthropic")
        
        elif self.provider == LLM_PROVIDER_OLLAMA:
            self.ollama_base_url = os.environ.get("OLLAMA_BASE_URL", "http://localhost:11434")
            self.ollama_model = os.environ.get("OLLAMA_MODEL", "qwen2.5:7b")
            print(f"✅ Ollama configured (base_url: {self.ollama_base_url}, model: {self.ollama_model})")
        
        elif self.provider == LLM_PROVIDER_HUGGINGFACE:
            self.hf_api_key = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACE_API_KEY")
            self.hf_model = os.environ.get("HF_MODEL", "Qwen/Qwen2.5-7B-Instruct")
            if self.hf_api_key:
                print(f"✅ Hugging Face API configured (model: {self.hf_model})")
            else:
                print("⚠️ HF_TOKEN not found, Hugging Face may have rate limits")
        
        elif self.provider == LLM_PROVIDER_API:
            # API mode - call HF Spaces API
            self.api_base_url = os.environ.get(
                "HF_API_BASE_URL", 
                "https://davidtran999-hue-portal-backend.hf.space/api"
            )
            print(f"✅ API mode configured (base_url: {self.api_base_url})")
        
        elif self.provider == LLM_PROVIDER_LOCAL:
            self._initialize_local_model()
        
        else:
            print("ℹ️ No LLM provider configured, using template-based generation")
    
    def _initialize_local_model(self):
        """Initialize local Hugging Face Transformers model."""
        try:
            from transformers import AutoModelForCausalLM, AutoTokenizer
            import torch
            
            # Default to Qwen 2.5 7B with 8-bit quantization (fits in GPU RAM)
            model_path = os.environ.get("LOCAL_MODEL_PATH", "Qwen/Qwen2.5-7B-Instruct")
            device = os.environ.get("LOCAL_MODEL_DEVICE", "auto")  # auto, cpu, cuda
            
            print(f"[LLM] Loading local model: {model_path}", flush=True)
            logger.info(f"[LLM] Loading local model: {model_path}")
            
            # Determine device
            if device == "auto":
                device = "cuda" if torch.cuda.is_available() else "cpu"
            
            # Start cache monitoring for download progress (optional)
            try:
                from .cache_monitor import get_cache_monitor
                monitor = get_cache_monitor()
                monitor.start_monitoring(model_path, interval=2.0)
                print(f"[LLM] 📊 Started cache monitoring for {model_path}", flush=True)
                logger.info(f"[LLM] 📊 Started cache monitoring for {model_path}")
            except Exception as e:
                logger.warning(f"Could not start cache monitoring: {e}")
            
            # Load tokenizer
            print("[LLM] Loading tokenizer...", flush=True)
            logger.info("[LLM] Loading tokenizer...")
            try:
                self.local_tokenizer = AutoTokenizer.from_pretrained(
                    model_path,
                    trust_remote_code=True
                )
                print("[LLM] ✅ Tokenizer loaded successfully", flush=True)
                logger.info("[LLM] ✅ Tokenizer loaded successfully")
            except Exception as tokenizer_err:
                error_trace = traceback.format_exc()
                print(f"[LLM] ❌ Tokenizer load error: {tokenizer_err}", flush=True)
                print(f"[LLM] ❌ Tokenizer trace: {error_trace}", flush=True)
                logger.error(f"[LLM] ❌ Tokenizer load error: {tokenizer_err}\n{error_trace}")
                print(f"[LLM] ❌ ERROR: {type(tokenizer_err).__name__}: {str(tokenizer_err)}", file=sys.stderr, flush=True)
                traceback.print_exc(file=sys.stderr)
                raise
            
            # Load model with optional quantization and fallback mechanism
            print(f"[LLM] Loading model to {device}...", flush=True)
            logger.info(f"[LLM] Loading model to {device}...")
            
            # Check for quantization config
            # Default to 8-bit for 7B (better thinking), 4-bit for larger models
            default_8bit = "7b" in model_path.lower() or "7B" in model_path
            default_4bit = ("32b" in model_path.lower() or "32B" in model_path or "14b" in model_path.lower() or "14B" in model_path) and not default_8bit
            
            # Check environment variable for explicit quantization preference
            quantization_pref = os.environ.get("LOCAL_MODEL_QUANTIZATION", "").lower()
            if quantization_pref == "4bit":
                use_8bit = False
                use_4bit = True
            elif quantization_pref == "8bit":
                use_8bit = True
                use_4bit = False
            elif quantization_pref == "none":
                use_8bit = False
                use_4bit = False
            else:
                # Use defaults based on model size
                use_8bit = os.environ.get("LOCAL_MODEL_8BIT", "true" if default_8bit else "false").lower() == "true"
                use_4bit = os.environ.get("LOCAL_MODEL_4BIT", "true" if default_4bit else "false").lower() == "true"
            
            # Try loading with fallback: 8-bit → 4-bit → float16
            model_loaded = False
            quantization_attempts = []
            
            if device == "cuda":
                # Attempt 1: Try 8-bit quantization (if requested)
                if use_8bit:
                    quantization_attempts.append(("8-bit", True, False))
                
                # Attempt 2: Try 4-bit quantization (if 8-bit fails or not requested)
                if use_4bit or (use_8bit and not model_loaded):
                    quantization_attempts.append(("4-bit", False, True))
                
                # Attempt 3: Fallback to float16 (no quantization)
                quantization_attempts.append(("float16", False, False))
            else:
                # CPU: only float32
                quantization_attempts.append(("float32", False, False))
            
            last_error = None
            for attempt_name, try_8bit, try_4bit in quantization_attempts:
                if model_loaded:
                    break
                
                try:
                    load_kwargs = {
                        "trust_remote_code": True,
                        "low_cpu_mem_usage": True,
                    }
                    
                    if device == "cuda":
                        load_kwargs["device_map"] = "auto"
                        
                        if try_4bit:
                            from transformers import BitsAndBytesConfig
                            load_kwargs["quantization_config"] = BitsAndBytesConfig(
                                load_in_4bit=True,
                                bnb_4bit_compute_dtype=torch.float16
                            )
                            print(f"[LLM] Attempting to load with 4-bit quantization (~4-5GB VRAM for 7B)", flush=True)
                        elif try_8bit:
                            from transformers import BitsAndBytesConfig
                            # Fixed: Remove CPU offload to avoid Int8Params compatibility issue
                            load_kwargs["quantization_config"] = BitsAndBytesConfig(
                                load_in_8bit=True,
                                llm_int8_threshold=6.0
                                # Removed: llm_int8_enable_fp32_cpu_offload=True (causes compatibility issues)
                            )
                            # Removed: max_memory override - let accelerate handle it automatically
                            print(f"[LLM] Attempting to load with 8-bit quantization (~7GB VRAM for 7B)", flush=True)
                        else:
                            load_kwargs["torch_dtype"] = torch.float16
                            print(f"[LLM] Attempting to load with float16 (no quantization)", flush=True)
                    else:
                        load_kwargs["torch_dtype"] = torch.float32
                        print(f"[LLM] Attempting to load with float32 (CPU)", flush=True)
                    
                    # Load model
                    self.local_model = AutoModelForCausalLM.from_pretrained(
                        model_path,
                        **load_kwargs
                    )
                    
                    # Stop cache monitoring (download complete)
                    try:
                        from .cache_monitor import get_cache_monitor
                        monitor = get_cache_monitor()
                        monitor.stop_monitoring(model_path)
                        print(f"[LLM] ✅ Model download complete, stopped monitoring", flush=True)
                    except:
                        pass
                    
                    print(f"[LLM] ✅ Model loaded successfully with {attempt_name} quantization", flush=True)
                    logger.info(f"[LLM] ✅ Model loaded successfully with {attempt_name} quantization")
                    model_loaded = True
                    
                except Exception as model_load_err:
                    last_error = model_load_err
                    error_trace = traceback.format_exc()
                    print(f"[LLM] ⚠️ Failed to load with {attempt_name}: {model_load_err}", flush=True)
                    logger.warning(f"[LLM] ⚠️ Failed to load with {attempt_name}: {model_load_err}")
                    
                    # If this was the last attempt, raise the error
                    if attempt_name == quantization_attempts[-1][0]:
                        print(f"[LLM] ❌ All quantization attempts failed. Last error: {model_load_err}", flush=True)
                        print(f"[LLM] ❌ Model load trace: {error_trace}", flush=True)
                        logger.error(f"[LLM] ❌ Model load error: {model_load_err}\n{error_trace}")
                        print(f"[LLM] ❌ ERROR: {type(model_load_err).__name__}: {str(model_load_err)}", file=sys.stderr, flush=True)
                        traceback.print_exc(file=sys.stderr)
                        raise
                    else:
                        # Try next quantization method
                        print(f"[LLM] 🔄 Falling back to next quantization method...", flush=True)
                        continue
            
            if not model_loaded:
                raise RuntimeError("Failed to load model with any quantization method")
            
            if device == "cpu":
                try:
                    self.local_model = self.local_model.to(device)
                    print(f"[LLM] ✅ Model moved to {device}", flush=True)
                    logger.info(f"[LLM] ✅ Model moved to {device}")
                except Exception as move_err:
                    error_trace = traceback.format_exc()
                    print(f"[LLM] ❌ Model move error: {move_err}", flush=True)
                    logger.error(f"[LLM] ❌ Model move error: {move_err}\n{error_trace}")
                    print(f"[LLM] ❌ ERROR: {type(move_err).__name__}: {str(move_err)}", file=sys.stderr, flush=True)
                    traceback.print_exc(file=sys.stderr)
            
            self.local_model.eval()  # Set to evaluation mode
            print(f"[LLM] ✅ Local model loaded successfully on {device}", flush=True)
            logger.info(f"[LLM] ✅ Local model loaded successfully on {device}")
            
        except ImportError as import_err:
            error_msg = "transformers package not installed, install with: pip install transformers torch"
            print(f"[LLM] ⚠️ {error_msg}", flush=True)
            logger.warning(f"[LLM] ⚠️ {error_msg}")
            print(f"[LLM] ❌ ImportError: {import_err}", file=sys.stderr, flush=True)
            self.local_model = None
            self.local_tokenizer = None
        except Exception as e:
            error_trace = traceback.format_exc()
            print(f"[LLM] ❌ Error loading local model: {e}", flush=True)
            print(f"[LLM] ❌ Full trace: {error_trace}", flush=True)
            logger.error(f"[LLM] ❌ Error loading local model: {e}\n{error_trace}")
            print(f"[LLM] ❌ ERROR: {type(e).__name__}: {str(e)}", file=sys.stderr, flush=True)
            traceback.print_exc(file=sys.stderr)
            print("[LLM] 💡 Tip: Use smaller models like Qwen/Qwen2.5-1.5B-Instruct or Qwen/Qwen2.5-0.5B-Instruct", flush=True)
            self.local_model = None
            self.local_tokenizer = None
    
    def is_available(self) -> bool:
        """Check if LLM is available."""
        return (
            self.client is not None or 
            self.provider == LLM_PROVIDER_OLLAMA or
            self.provider == LLM_PROVIDER_HUGGINGFACE or
            self.provider == LLM_PROVIDER_API or
            (self.provider == LLM_PROVIDER_LOCAL and self.local_model is not None)
        )
    
    def generate_answer(
        self,
        query: str,
        context: Optional[List[Dict[str, Any]]] = None,
        documents: Optional[List[Any]] = None
    ) -> Optional[str]:
        """
        Generate natural language answer from documents.
        
        Args:
            query: User query.
            context: Optional conversation context.
            documents: Retrieved documents.
        
        Returns:
            Generated answer or None if LLM not available.
        """
        if not self.is_available():
            return None
        
        # Build prompt
        prompt = self._build_prompt(query, context, documents)
        
        try:
            print(f"[LLM] Generating answer with provider: {self.provider}", flush=True)
            logger.info(f"[LLM] Generating answer with provider: {self.provider}")
            
            if self.provider == LLM_PROVIDER_OPENAI:
                result = self._generate_openai(prompt)
            elif self.provider == LLM_PROVIDER_ANTHROPIC:
                result = self._generate_anthropic(prompt)
            elif self.provider == LLM_PROVIDER_OLLAMA:
                result = self._generate_ollama(prompt)
            elif self.provider == LLM_PROVIDER_HUGGINGFACE:
                result = self._generate_huggingface(prompt)
            elif self.provider == LLM_PROVIDER_LOCAL:
                result = self._generate_local(prompt)
            elif self.provider == LLM_PROVIDER_API:
                # For API mode, send the full prompt (with documents) as the message
                # This ensures HF Spaces receives all context from retrieved documents
                result = self._generate_api(prompt, context)
            else:
                result = None
            
            if result:
                print(f"[LLM] ✅ Answer generated successfully (length: {len(result)})", flush=True)
                logger.info(f"[LLM] ✅ Answer generated successfully (length: {len(result)})")
            else:
                print(f"[LLM] ⚠️ No answer generated", flush=True)
                logger.warning("[LLM] ⚠️ No answer generated")
            
            return result
        except Exception as e:
            error_trace = traceback.format_exc()
            print(f"[LLM] ❌ Error generating answer: {e}", flush=True)
            print(f"[LLM] ❌ Full trace: {error_trace}", flush=True)
            logger.error(f"[LLM] ❌ Error generating answer: {e}\n{error_trace}")
            print(f"[LLM] ❌ ERROR: {type(e).__name__}: {str(e)}", file=sys.stderr, flush=True)
            traceback.print_exc(file=sys.stderr)
            return None
    
    def _build_prompt(
        self,
        query: str,
        context: Optional[List[Dict[str, Any]]],
        documents: Optional[List[Any]]
    ) -> str:
        """Build prompt for LLM."""
        prompt_parts = [
            "Bạn là chatbot tư vấn pháp lý của Công an Thừa Thiên Huế.",
            "Nhiệm vụ: Trả lời câu hỏi của người dùng dựa trên các văn bản pháp luật và quy định được cung cấp.",
            "",
            f"Câu hỏi của người dùng: {query}",
            ""
        ]
        
        if context:
            prompt_parts.append("Ngữ cảnh cuộc hội thoại trước đó:")
            for msg in context[-3:]:  # Last 3 messages
                role = "Người dùng" if msg.get("role") == "user" else "Bot"
                content = msg.get("content", "")
                prompt_parts.append(f"{role}: {content}")
            prompt_parts.append("")
        
        if documents:
            prompt_parts.append("Các văn bản/quy định liên quan:")
            for i, doc in enumerate(documents[:5], 1):
                # Extract relevant fields based on document type
                doc_text = self._format_document(doc)
                prompt_parts.append(f"{i}. {doc_text}")
            prompt_parts.append("")
            # If documents exist, require strict adherence
            prompt_parts.extend([
                "Yêu cầu QUAN TRỌNG:",
                "- CHỈ trả lời dựa trên thông tin trong 'Các văn bản/quy định liên quan' ở trên",
                "- KHÔNG được tự tạo hoặc suy đoán thông tin không có trong tài liệu",
                "- Nếu thông tin không đủ để trả lời, hãy nói rõ: 'Thông tin trong cơ sở dữ liệu chưa đủ để trả lời câu hỏi này'",
                "- Nếu có mức phạt, phải ghi rõ số tiền (ví dụ: 200.000 - 400.000 VNĐ)",
                "- Nếu có điều khoản, ghi rõ mã điều (ví dụ: Điều 5, Điều 10)",
                "- Nếu có thủ tục, ghi rõ hồ sơ, lệ phí, thời hạn",
                "- Trả lời bằng tiếng Việt, ngắn gọn, dễ hiểu",
                "",
                "Trả lời:"
            ])
        else:
            # No documents - allow general conversation
            prompt_parts.extend([
                "Yêu cầu:",
                "- Trả lời câu hỏi một cách tự nhiên và hữu ích như một chatbot AI thông thường",
                "- Nếu câu hỏi liên quan đến pháp luật, thủ tục, mức phạt nhưng không có thông tin trong cơ sở dữ liệu, hãy nói: 'Tôi không tìm thấy thông tin này trong cơ sở dữ liệu. Bạn có thể liên hệ trực tiếp với Công an Thừa Thiên Huế để được tư vấn chi tiết hơn.'",
                "- Trả lời bằng tiếng Việt, thân thiện, ngắn gọn, dễ hiểu",
                "",
                "Trả lời:"
            ])
        
        return "\n".join(prompt_parts)
    
    def _format_document(self, doc: Any) -> str:
        """Format document for prompt."""
        doc_type = type(doc).__name__.lower()
        
        if "fine" in doc_type:
            parts = [f"Mức phạt: {getattr(doc, 'name', '')}"]
            if hasattr(doc, 'code') and doc.code:
                parts.append(f"Mã: {doc.code}")
            if hasattr(doc, 'min_fine') and hasattr(doc, 'max_fine'):
                if doc.min_fine and doc.max_fine:
                    parts.append(f"Số tiền: {doc.min_fine:,.0f} - {doc.max_fine:,.0f} VNĐ")
            return " | ".join(parts)
        
        elif "procedure" in doc_type:
            parts = [f"Thủ tục: {getattr(doc, 'title', '')}"]
            if hasattr(doc, 'dossier') and doc.dossier:
                parts.append(f"Hồ sơ: {doc.dossier}")
            if hasattr(doc, 'fee') and doc.fee:
                parts.append(f"Lệ phí: {doc.fee}")
            return " | ".join(parts)
        
        elif "office" in doc_type:
            parts = [f"Đơn vị: {getattr(doc, 'unit_name', '')}"]
            if hasattr(doc, 'address') and doc.address:
                parts.append(f"Địa chỉ: {doc.address}")
            if hasattr(doc, 'phone') and doc.phone:
                parts.append(f"Điện thoại: {doc.phone}")
            return " | ".join(parts)
        
        elif "advisory" in doc_type:
            parts = [f"Cảnh báo: {getattr(doc, 'title', '')}"]
            if hasattr(doc, 'summary') and doc.summary:
                parts.append(f"Nội dung: {doc.summary[:200]}")
            return " | ".join(parts)
        
        elif "legalsection" in doc_type or "legal" in doc_type:
            parts = []
            if hasattr(doc, 'section_code') and doc.section_code:
                parts.append(f"Điều khoản: {doc.section_code}")
            if hasattr(doc, 'section_title') and doc.section_title:
                parts.append(f"Tiêu đề: {doc.section_title}")
            if hasattr(doc, 'document') and doc.document:
                doc_obj = doc.document
                if hasattr(doc_obj, 'title'):
                    parts.append(f"Văn bản: {doc_obj.title}")
                if hasattr(doc_obj, 'code'):
                    parts.append(f"Mã văn bản: {doc_obj.code}")
            if hasattr(doc, 'content') and doc.content:
                # Truncate content to 300 chars for prompt
                content_short = doc.content[:300] + "..." if len(doc.content) > 300 else doc.content
                parts.append(f"Nội dung: {content_short}")
            return " | ".join(parts) if parts else str(doc)
        
        return str(doc)
    
    def _generate_openai(self, prompt: str) -> Optional[str]:
        """Generate answer using OpenAI."""
        if not self.client:
            return None
        
        try:
            response = self.client.chat.completions.create(
                model=os.environ.get("OPENAI_MODEL", "gpt-3.5-turbo"),
                messages=[
                    {"role": "system", "content": "Bạn là chatbot tư vấn chuyên nghiệp."},
                    {"role": "user", "content": prompt}
                ],
                temperature=0.7,
                max_tokens=500
            )
            return response.choices[0].message.content
        except Exception as e:
            print(f"OpenAI API error: {e}")
            return None
    
    def _generate_anthropic(self, prompt: str) -> Optional[str]:
        """Generate answer using Anthropic Claude."""
        if not self.client:
            return None
        
        try:
            message = self.client.messages.create(
                model=os.environ.get("ANTHROPIC_MODEL", "claude-3-5-sonnet-20241022"),
                max_tokens=500,
                messages=[
                    {"role": "user", "content": prompt}
                ]
            )
            return message.content[0].text
        except Exception as e:
            print(f"Anthropic API error: {e}")
            return None
    
    def _generate_ollama(self, prompt: str) -> Optional[str]:
        """Generate answer using Ollama (local LLM)."""
        try:
            import requests
            model = getattr(self, 'ollama_model', os.environ.get("OLLAMA_MODEL", "qwen2.5:7b"))
            
            response = requests.post(
                f"{self.ollama_base_url}/api/generate",
                json={
                    "model": model,
                    "prompt": prompt,
                    "stream": False,
                    "options": {
                        "temperature": 0.7,
                        "top_p": 0.9,
                        "num_predict": 500
                    }
                },
                timeout=60
            )
            
            if response.status_code == 200:
                return response.json().get("response")
            return None
        except Exception as e:
            print(f"Ollama API error: {e}")
            return None
    
    def _generate_huggingface(self, prompt: str) -> Optional[str]:
        """Generate answer using Hugging Face Inference API."""
        try:
            import requests
            
            api_url = f"https://api-inference.huggingface.co/models/{self.hf_model}"
            headers = {}
            if hasattr(self, 'hf_api_key') and self.hf_api_key:
                headers["Authorization"] = f"Bearer {self.hf_api_key}"
            
            response = requests.post(
                api_url,
                headers=headers,
                json={
                    "inputs": prompt,
                    "parameters": {
                        "temperature": 0.7,
                        "max_new_tokens": 500,
                        "return_full_text": False
                    }
                },
                timeout=60
            )
            
            if response.status_code == 200:
                result = response.json()
                if isinstance(result, list) and len(result) > 0:
                    return result[0].get("generated_text", "")
                elif isinstance(result, dict):
                    return result.get("generated_text", "")
            elif response.status_code == 503:
                # Model is loading, wait and retry
                print("⚠️ Model is loading, please wait...")
                return None
            else:
                print(f"Hugging Face API error: {response.status_code} - {response.text}")
            return None
        except Exception as e:
            print(f"Hugging Face API error: {e}")
            return None
    
    def _generate_local(self, prompt: str) -> Optional[str]:
        """Generate answer using local Hugging Face Transformers model."""
        if self.local_model is None or self.local_tokenizer is None:
            return None
        
        try:
            import torch
            
            # Format prompt for Qwen models
            messages = [
                {"role": "system", "content": "Bạn là chatbot tư vấn chuyên nghiệp."},
                {"role": "user", "content": prompt}
            ]
            
            # Apply chat template if available
            if hasattr(self.local_tokenizer, "apply_chat_template"):
                text = self.local_tokenizer.apply_chat_template(
                    messages,
                    tokenize=False,
                    add_generation_prompt=True
                )
            else:
                text = prompt
            
            # Tokenize
            inputs = self.local_tokenizer(text, return_tensors="pt")
            
            # Move to device
            device = next(self.local_model.parameters()).device
            inputs = {k: v.to(device) for k, v in inputs.items()}
            
            # Generate
            with torch.no_grad():
                outputs = self.local_model.generate(
                    **inputs,
                    max_new_tokens=500,
                    temperature=0.7,
                    top_p=0.9,
                    do_sample=True,
                    pad_token_id=self.local_tokenizer.eos_token_id
                )
            
            # Decode
            generated_text = self.local_tokenizer.decode(
                outputs[0][inputs["input_ids"].shape[1]:],
                skip_special_tokens=True
            )
            
            return generated_text.strip()
            
        except TypeError as e:
            # Check for Int8Params compatibility error
            if "_is_hf_initialized" in str(e) or "Int8Params" in str(e):
                error_msg = (
                    f"[LLM] ❌ Int8Params compatibility error: {e}\n"
                    f"[LLM] 💡 This error occurs when using 8-bit quantization with incompatible library versions.\n"
                    f"[LLM] 💡 Solutions:\n"
                    f"[LLM]   1. Set LOCAL_MODEL_QUANTIZATION=4bit to use 4-bit quantization instead\n"
                    f"[LLM]   2. Set LOCAL_MODEL_QUANTIZATION=none to disable quantization\n"
                    f"[LLM]   3. Use API mode (LLM_PROVIDER=api) to avoid local model issues\n"
                    f"[LLM]   4. Use a smaller model like Qwen/Qwen2.5-1.5B-Instruct"
                )
                print(error_msg, flush=True)
                logger.error(f"[LLM] ❌ Int8Params compatibility error: {e}")
                print(f"[LLM] ❌ ERROR: {type(e).__name__}: {str(e)}", file=sys.stderr, flush=True)
                return None
            else:
                # Other TypeError, re-raise to be caught by general handler
                raise
        except Exception as e:
            error_trace = traceback.format_exc()
            print(f"[LLM] ❌ Local model generation error: {e}", flush=True)
            print(f"[LLM] ❌ Full trace: {error_trace}", flush=True)
            logger.error(f"[LLM] ❌ Local model generation error: {e}\n{error_trace}")
            print(f"[LLM] ❌ ERROR: {type(e).__name__}: {str(e)}", file=sys.stderr, flush=True)
            traceback.print_exc(file=sys.stderr)
            return None
    
    def _generate_api(self, prompt: str, context: Optional[List[Dict[str, Any]]] = None) -> Optional[str]:
        """Generate answer by calling HF Spaces API.
        
        Args:
            prompt: Full prompt including query and documents context.
            context: Optional conversation context (not used in API mode, handled by HF Spaces).
        """
        if not self.api_base_url:
            return None
        
        try:
            import requests
            
            # Prepare request payload
            # Send the full prompt (with documents) as the message to HF Spaces
            # This ensures HF Spaces receives all context from retrieved documents
            payload = {
                "message": prompt,
                "reset_session": False
            }
            
            # Only add session_id if we have a valid session context
            # For now, we'll omit it and let the API generate a new one
            
            # Add context if available (API may support this in future)
            # For now, context is handled by the API internally
            
            # Call API endpoint
            api_url = f"{self.api_base_url}/chatbot/chat/"
            print(f"[LLM] 🔗 Calling API: {api_url}", flush=True)
            print(f"[LLM] 📤 Payload: {payload}", flush=True)
            
            response = requests.post(
                api_url,
                json=payload,
                headers={"Content-Type": "application/json"},
                timeout=60
            )
            
            print(f"[LLM] 📥 Response status: {response.status_code}", flush=True)
            print(f"[LLM] 📥 Response headers: {dict(response.headers)}", flush=True)
            
            if response.status_code == 200:
                try:
                    result = response.json()
                    print(f"[LLM] 📥 Response JSON: {result}", flush=True)
                    # Extract message from response
                    if isinstance(result, dict):
                        message = result.get("message", None)
                        if message:
                            print(f"[LLM] ✅ Got message from API (length: {len(message)})", flush=True)
                        return message
                    else:
                        print(f"[LLM] ⚠️ Response is not a dict: {type(result)}", flush=True)
                        return None
                except ValueError as e:
                    print(f"[LLM] ❌ JSON decode error: {e}", flush=True)
                    print(f"[LLM] ❌ Response text: {response.text[:500]}", flush=True)
                    return None
            elif response.status_code == 503:
                # Service unavailable - model might be loading
                print("[LLM] ⚠️ API service is loading, please wait...", flush=True)
                return None
            else:
                print(f"[LLM] ❌ API error: {response.status_code} - {response.text[:500]}", flush=True)
                return None
        except requests.exceptions.Timeout:
            print("[LLM] ❌ API request timeout")
            return None
        except requests.exceptions.ConnectionError as e:
            print(f"[LLM] ❌ API connection error: {e}")
            return None
        except Exception as e:
            error_trace = traceback.format_exc()
            print(f"[LLM] ❌ API mode error: {e}", flush=True)
            print(f"[LLM] ❌ Full trace: {error_trace}", flush=True)
            logger.error(f"[LLM] ❌ API mode error: {e}\n{error_trace}")
            return None
    
    def summarize_context(self, messages: List[Dict[str, Any]], max_length: int = 200) -> str:
        """
        Summarize conversation context.
        
        Args:
            messages: List of conversation messages.
            max_length: Maximum summary length.
        
        Returns:
            Summary string.
        """
        if not messages:
            return ""
        
        # Simple summarization: extract key entities and intents
        intents = []
        entities = set()
        
        for msg in messages:
            if msg.get("intent"):
                intents.append(msg["intent"])
            if msg.get("entities"):
                for key, value in msg["entities"].items():
                    if isinstance(value, str):
                        entities.add(value)
                    elif isinstance(value, list):
                        entities.update(value)
        
        summary_parts = []
        if intents:
            unique_intents = list(set(intents))
            summary_parts.append(f"Chủ đề: {', '.join(unique_intents)}")
        if entities:
            summary_parts.append(f"Thông tin: {', '.join(list(entities)[:5])}")
        
        summary = ". ".join(summary_parts)
        return summary[:max_length] if len(summary) > max_length else summary
    
    def extract_entities_llm(self, query: str) -> Dict[str, Any]:
        """
        Extract entities using LLM.
        
        Args:
            query: User query.
        
        Returns:
            Dictionary of extracted entities.
        """
        if not self.is_available():
            return {}
        
        prompt = f"""
        Trích xuất các thực thể từ câu hỏi sau:
        "{query}"
        
        Các loại thực thể cần tìm:
        - fine_code: Mã vi phạm (V001, V002, ...)
        - fine_name: Tên vi phạm
        - procedure_name: Tên thủ tục
        - office_name: Tên đơn vị
        
        Trả lời dưới dạng JSON: {{"fine_code": "...", "fine_name": "...", ...}}
        Nếu không có, trả về {{}}.
        """
        
        try:
            if self.provider == LLM_PROVIDER_OPENAI:
                response = self._generate_openai(prompt)
            elif self.provider == LLM_PROVIDER_ANTHROPIC:
                response = self._generate_anthropic(prompt)
            elif self.provider == LLM_PROVIDER_OLLAMA:
                response = self._generate_ollama(prompt)
            elif self.provider == LLM_PROVIDER_HUGGINGFACE:
                response = self._generate_huggingface(prompt)
            elif self.provider == LLM_PROVIDER_LOCAL:
                response = self._generate_local(prompt)
            elif self.provider == LLM_PROVIDER_API:
                # For API mode, we can't extract entities directly
                # Return empty dict
                return {}
            else:
                return {}
            
            if response:
                # Try to extract JSON from response
                json_match = re.search(r'\{[^}]+\}', response)
                if json_match:
                    return json.loads(json_match.group())
        except Exception as e:
            print(f"Error extracting entities with LLM: {e}")
        
        return {}


# Global LLM generator instance
_llm_generator: Optional[LLMGenerator] = None
_last_provider: Optional[str] = None

def get_llm_generator() -> Optional[LLMGenerator]:
    """Get or create LLM generator instance.
    
    Recreates instance if provider changed (e.g., from local to api).
    """
    global _llm_generator, _last_provider
    
    # Get current provider from env
    current_provider = os.environ.get("LLM_PROVIDER", LLM_PROVIDER_NONE).lower()
    
    # Recreate if provider changed or instance doesn't exist
    if _llm_generator is None or _last_provider != current_provider:
        _llm_generator = LLMGenerator()
        _last_provider = current_provider
        print(f"[LLM] 🔄 Recreated LLM generator with provider: {current_provider}", flush=True)
    
    return _llm_generator if _llm_generator.is_available() else None