File size: 19,736 Bytes
401184c
 
a93e2b1
401184c
 
 
 
 
 
 
 
 
 
 
 
 
a93e2b1
 
 
401184c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
38d12c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
401184c
 
 
 
 
 
 
 
 
 
 
 
b764fd1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
401184c
 
 
 
 
 
 
8821b53
401184c
8821b53
401184c
 
 
8821b53
401184c
 
 
 
 
 
 
 
 
 
 
c3f88f2
401184c
 
 
 
 
 
 
 
8821b53
 
 
 
 
 
 
 
 
 
 
 
 
401184c
 
c3f88f2
 
 
 
 
 
 
 
 
 
401184c
 
 
 
 
8821b53
401184c
 
 
 
8821b53
 
 
 
 
 
 
 
 
 
 
 
401184c
8821b53
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b764fd1
401184c
8821b53
 
 
 
 
 
 
 
 
 
c6e3d43
 
 
 
 
 
8821b53
c6e3d43
8821b53
 
 
401184c
 
 
 
 
 
 
 
38d12c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
401184c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8f03f27
 
 
401184c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8f03f27
 
 
 
401184c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
543b684
401184c
 
 
 
543b684
401184c
 
 
 
 
 
 
 
 
a93e2b1
401184c
 
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
"""
专业幻觉检测模块
支持多种检测方法:NLI模型、专门检测模型、轻量级模型、混合检测
"""

import re
from typing import List, Dict, Tuple
import torch
from transformers import (
    AutoModelForSequenceClassification,
    AutoTokenizer,
    pipeline
)
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np

# 导入轻量级检测器
from lightweight_hallucination_detector import LightweightHallucinationDetector


class VectaraHallucinationDetector:
    """
    Vectara 专门的幻觉检测模型
    使用 HHEM (Hughes Hallucination Evaluation Model)
    """
    
    def __init__(self):
        """初始化 Vectara 幻觉检测模型"""
        print("🔧 初始化 Vectara 幻觉检测模型...")
        
        try:
            self.model_name = "vectara/hallucination_evaluation_model"
            self.tokenizer = AutoTokenizer.from_pretrained(self.model_name)
            self.model = AutoModelForSequenceClassification.from_pretrained(self.model_name)
            self.model.eval()  # 设置为评估模式
            
            # 移动到GPU(如果可用)
            self.device = "cuda" if torch.cuda.is_available() else "cpu"
            self.model.to(self.device)
            
            print(f"✅ Vectara 模型加载成功 (device: {self.device})")
        except Exception as e:
            print(f"⚠️ Vectara 模型加载失败: {e}")
            print("💡 尝试使用 NLI 模型作为备选...")
            self.model = None
    
    def detect(self, generation: str, documents: str) -> Dict:
        """
        检测幻觉
        
        Args:
            generation: LLM 生成的内容
            documents: 参考文档
            
        Returns:
            {
                "has_hallucination": bool,
                "hallucination_score": float (0-1),
                "factuality_score": float (0-1)
            }
        """
        if self.model is None:
            return {"has_hallucination": False, "hallucination_score": 0.0, "factuality_score": 1.0}
        
        try:
            # 准备输入
            inputs = self.tokenizer(
                documents,
                generation,
                return_tensors="pt",
                truncation=True,
                max_length=512,
                padding=True
            ).to(self.device)
            
            # 推理
            with torch.no_grad():
                outputs = self.model(**inputs)
                logits = outputs.logits
                probs = torch.softmax(logits, dim=-1)
            
            # Vectara 模型输出:[0] = factual, [1] = hallucinated
            factuality_score = probs[0][0].item()
            hallucination_score = probs[0][1].item()
            
            # 判断是否有幻觉(阈值 0.5)
            has_hallucination = hallucination_score > 0.5
            
            return {
                "has_hallucination": has_hallucination,
                "hallucination_score": hallucination_score,
                "factuality_score": factuality_score
            }
        
        except Exception as e:
            print(f"❌ Vectara 检测失败: {e}")
            return {"has_hallucination": False, "hallucination_score": 0.0, "factuality_score": 1.0}
    
    def grade(self, generation: str, documents) -> str:
        """
        兼容原有接口的检测方法
        
        Args:
            generation: LLM 生成的内容
            documents: 参考文档(可以是字符串或列表)
            
        Returns:
            "yes" 表示无幻觉,"no" 表示有幻觉
        """
        # 处理文档格式
        if isinstance(documents, list):
            doc_text = "\n\n".join([
                doc.page_content if hasattr(doc, 'page_content') else str(doc)
                for doc in documents
            ])
        else:
            doc_text = str(documents)
        
        # 检测幻觉
        result = self.detect(generation, doc_text)
        
        # 打印详细信息
        if result['has_hallucination']:
            print(f"⚠️ Vectara 检测到幻觉 (得分: {result['hallucination_score']:.2f})")
        else:
            print(f"✅ Vectara 未检测到幻觉 (真实性得分: {result['factuality_score']:.2f})")
        
        # 返回兼容格式
        return "no" if result['has_hallucination'] else "yes"


class NLIHallucinationDetector:
    """
    基于 NLI (Natural Language Inference) 的幻觉检测
    使用 DeBERTa 模型
    """
    
    def __init__(self):
        """初始化 NLI 模型"""
        print("🔧 初始化 NLI 幻觉检测模型...")
        
        # 尝试多个模型,按照从小到大的顺序
        models_to_try = [
            "cross-encoder/nli-deberta-v3-xsmall",  # 最小 40MB
            "cross-encoder/nli-deberta-v3-small",   # 小 150MB
            "cross-encoder/nli-MiniLM2-L6-H768",    # 轻量 90MB
            "facebook/bart-large-mnli",              # 备用
        ]
        
        self.nli_model = None
        
        for model_name in models_to_try:
            try:
                print(f"   尝试加载: {model_name}...")
                self.nli_model = pipeline(
                    "text-classification",
                    model=model_name,
                    device=0 if torch.cuda.is_available() else -1,
                    truncation=True,
                    max_length=512
                )
                print(f"✅ NLI 模型加载成功: {model_name}")
                self.model_name = model_name
                break  # 成功加载,退出循环
            except Exception as e:
                print(f"   ⚠️ {model_name} 加载失败: {str(e)[:80]}")
                continue
        
        if self.nli_model is None:
            print("❌ 所有 NLI 模型加载失败,将禁用 NLI 检测")
    
    def split_sentences(self, text: str) -> List[str]:
        """分割句子"""
        # 简单的句子分割(可以用更复杂的 NLP 工具)
        sentences = re.split(r'[。!?\.\!\?]\s*', text)
        return [s.strip() for s in sentences if s.strip()]
    
    def detect(self, generation: str, documents) -> Dict:
        """
        检测幻觉(支持多文档最大匹配策略)
        
        Args:
            generation: LLM 生成的内容
            documents: 参考文档 (str 或 List[Document/str])
            
        Returns:
            {
                "has_hallucination": bool,
                "contradiction_count": int,
                "neutral_count": int,
                "entailment_count": int,
                "problematic_sentences": List[str]
            }
        """
        if self.nli_model is None:
            print("⚠️ NLI 模型未加载,跳过检测")
            return {
                "has_hallucination": False,
                "contradiction_count": 0,
                "neutral_count": 0,
                "entailment_count": 0,
                "problematic_sentences": []
            }
        
        # 1. 预处理文档列表
        docs_content = []
        if isinstance(documents, list):
            for doc in documents:
                if hasattr(doc, 'page_content'):
                    docs_content.append(doc.page_content)
                else:
                    docs_content.append(str(doc))
        else:
            # 如果是单个字符串,尝试按换行符分割,或者作为单文档处理
            docs_content = [str(documents)]

        # 2. 分割生成内容为句子
        sentences = self.split_sentences(generation)
        
        if not sentences:
            print("⚠️ 没有检测到有效句子")
            return {
                "has_hallucination": False,
                "contradiction_count": 0,
                "neutral_count": 0,
                "entailment_count": 0,
                "problematic_sentences": []
            }
        
        contradiction_count = 0
        neutral_count = 0
        entailment_count = 0
        problematic_sentences = []
        
        # 3. 逐句检测 (Max-Entailment Strategy)
        for sentence in sentences:
            if len(sentence) < 10:  # 跳过太短的句子
                continue
            
            # 默认为 Neutral (找不到支持)
            best_label = "neutral"
            best_score = 0.0
            
            # 遍历所有文档块,寻找最佳匹配
            # 只要有一个文档能 Entail (支持) 这个句子,就算通过
            sentence_supported = False
            
            for doc_content in docs_content:
                # 截断单个文档块以适应模型 (保留前 800 字符,通常足够覆盖 512 tokens)
                # 注意:这里是对单个文档块截断,而不是对所有文档拼接后截断
                premise = doc_content[:800]
                
                try:
                    # NLI 推理
                    if hasattr(self, 'model_name') and 'cross-encoder' in self.model_name:
                        result = self.nli_model(
                            f"{premise} [SEP] {sentence}",
                            truncation=True,
                            max_length=512
                        )
                    else:
                        result = self.nli_model(
                            sentence,
                            premise,
                            truncation=True,
                            max_length=512
                        )
                    
                    # 解析结果
                    if isinstance(result, list) and len(result) > 0:
                        current_label = result[0]['label'].lower()
                        current_score = result[0]['score']
                        
                        # 优先级逻辑:Entailment > Contradiction > Neutral
                        # 如果找到 Entailment,立即停止查找(已验证)
                        if 'entailment' in current_label or 'entail' in current_label:
                            best_label = "entailment"
                            sentence_supported = True
                            break
                        
                        # 如果是 Contradiction,记录下来,但继续找(也许其他文档能解释)
                        if 'contradiction' in current_label or 'contradict' in current_label:
                            # 只有当目前是 Neutral 时才更新为 Contradiction
                            # 这样防止 Contradiction 覆盖了潜在的 Entailment (虽然上面break了,但这逻辑保持严谨)
                            if best_label == "neutral":
                                best_label = "contradiction"
                                best_score = current_score
                                
                    else:
                        continue
                        
                except Exception as e:
                    print(f"⚠️ NLI 子任务失败: {str(e)[:50]}")
                    continue
            
            # 统计该句子的最终判定
            if best_label == "entailment":
                entailment_count += 1
            elif best_label == "contradiction":
                contradiction_count += 1
                problematic_sentences.append(sentence)
            else: # neutral
                neutral_count += 1
                
        # 4. 综合评分
        total_sentences = contradiction_count + neutral_count + entailment_count
        
        has_hallucination = False
        if total_sentences > 0:
            contradiction_ratio = contradiction_count / total_sentences
            neutral_ratio = neutral_count / total_sentences
            # 阈值判断
            has_hallucination = (contradiction_ratio > 0.3) or (neutral_ratio > 0.8)
            
            # Debug 信息
            print(f"📊 NLI 检测结果: Entail={entailment_count}, Contra={contradiction_count}, Neutral={neutral_count}")
        
        return {
            "has_hallucination": has_hallucination,
            "contradiction_count": contradiction_count,
            "neutral_count": neutral_count,
            "entailment_count": entailment_count,
            "problematic_sentences": problematic_sentences
        }
    
    def grade(self, generation: str, documents) -> str:
        """
        兼容原有接口的检测方法
        
        Args:
            generation: LLM 生成的内容
            documents: 参考文档(可以是字符串或列表)
            
        Returns:
            "yes" 表示无幻觉,"no" 表示有幻觉
        """
        # 处理文档格式
        if isinstance(documents, list):
            doc_text = "\n\n".join([
                doc.page_content if hasattr(doc, 'page_content') else str(doc)
                for doc in documents
            ])
        else:
            doc_text = str(documents)
        
        # 检测幻觉
        result = self.detect(generation, doc_text)
        
        # 打印详细信息
        if result['has_hallucination']:
            print(f"⚠️ NLI 检测到幻觉")
            print(f"   矛盾句子: {result['contradiction_count']}")
            print(f"   中立句子: {result['neutral_count']}")
            print(f"   蕴含句子: {result['entailment_count']}")
            if result['problematic_sentences']:
                print(f"   问题句子: {result['problematic_sentences'][:2]}")
        else:
            print(f"✅ NLI 未检测到幻觉")
        
        # 返回兼容格式
        return "no" if result['has_hallucination'] else "yes"


class HybridHallucinationDetector:
    """
    混合幻觉检测器
    结合 Vectara 模型和 NLI 模型,提供最佳检测效果
    """
    
    def __init__(self, use_vectara: bool = True, use_nli: bool = True):
        """
        初始化混合检测器
        
        Args:
            use_vectara: 是否使用 Vectara 模型
            use_nli: 是否使用 NLI 模型
        """
        self.detectors = {}
        
        if use_vectara:
            try:
                self.detectors['vectara'] = VectaraHallucinationDetector()
            except Exception as e:
                print(f"⚠️ Vectara 检测器初始化失败: {e}")
        
        if use_nli:
            try:
                self.detectors['nli'] = NLIHallucinationDetector()
            except Exception as e:
                print(f"⚠️ NLI 检测器初始化失败: {e}")
        
        if not self.detectors:
            raise RuntimeError("❌ 所有检测器初始化失败!")
        
        print(f"✅ 混合检测器就绪,已加载: {list(self.detectors.keys())}")
    
    def detect(self, generation: str, documents: str) -> Dict:
        """
        综合检测幻觉
        
        Returns:
            {
                "has_hallucination": bool,
                "confidence": float,
                "vectara_result": Dict,
                "nli_result": Dict,
                "method_used": str
            }
        """
        results = {
            "has_hallucination": False,
            "confidence": 0.0,
            "method_used": ""
        }
        
        # 1. 优先使用 Vectara(最准确)
        if 'vectara' in self.detectors:
            vectara_result = self.detectors['vectara'].detect(generation, documents)
            results['vectara_result'] = vectara_result
            
            if vectara_result['hallucination_score'] > 0.3:  # 降低阈值以提高灵敏度
                results['has_hallucination'] = True
                results['confidence'] = vectara_result['hallucination_score']
                results['method_used'] = 'vectara'
                return results
            else:
                # Vectara 未检测到幻觉,设置 method_used
                results['method_used'] = 'vectara'
        
        # 2. 如果 Vectara 不确定或不可用,使用 NLI 二次确认
        if 'nli' in self.detectors:
            nli_result = self.detectors['nli'].detect(generation, documents)
            results['nli_result'] = nli_result
            
            if nli_result['has_hallucination']:
                results['has_hallucination'] = True
                # 计算置信度
                total_sentences = (nli_result['contradiction_count'] + 
                                 nli_result['neutral_count'] + 
                                 nli_result['entailment_count'])
                if total_sentences > 0:
                    results['confidence'] = (nli_result['contradiction_count'] + 
                                           nli_result['neutral_count'] * 0.5) / total_sentences
                results['method_used'] = 'nli'
            else:
                # 未检测到幻觉,也要设置 method_used
                if not results['method_used']:  # 只有当前面没有设置时
                    results['method_used'] = 'nli'
        
        # 如果两个模型都有结果,投票决定
        if 'vectara_result' in results and 'nli_result' in results:
            vectara_vote = results['vectara_result']['has_hallucination']
            nli_vote = results['nli_result']['has_hallucination']
            
            if vectara_vote and nli_vote:
                results['has_hallucination'] = True
                results['confidence'] = min(
                    results.get('vectara_result', {}).get('hallucination_score', 0.5),
                    results.get('confidence', 0.5)
                )
                results['method_used'] = 'vectara+nli'
        
        return results
    
    def grade(self, generation: str, documents) -> str:
        """
        兼容原有接口的检测方法
        
        Args:
            generation: LLM 生成的内容
            documents: 参考文档(可以是字符串或列表)
            
        Returns:
            "yes" 表示无幻觉,"no" 表示有幻觉
        """
        # 处理文档格式
        if isinstance(documents, list):
            doc_text = "\n\n".join([
                doc.page_content if hasattr(doc, 'page_content') else str(doc)
                for doc in documents
            ])
        else:
            doc_text = str(documents)
        
        # 检测幻觉
        result = self.detect(generation, doc_text)
        
        # 打印详细信息
        if result['has_hallucination']:
            print(f"⚠️ 检测到幻觉 (置信度: {result['confidence']:.2f}, 方法: {result['method_used']})")
            if 'nli_result' in result:
                print(f"   矛盾句子: {result['nli_result']['contradiction_count']}")
                if result['nli_result']['problematic_sentences']:
                    print(f"   问题句子: {result['nli_result']['problematic_sentences'][:2]}")
        else:
            print(f"✅ 未检测到幻觉 (方法: {result['method_used']})")
        
        # 返回兼容格式
        return "no" if result['has_hallucination'] else "yes"


def initialize_hallucination_detector(method: str = "nli") -> object:
    """
    初始化幻觉检测器
    
    Args:
        method: 'vectara', 'nli', 或 'hybrid' (推荐)
        
    Returns:
        幻觉检测器实例
    """
    if method == "vectara":
        return VectaraHallucinationDetector()
    elif method == "nli":
        return NLIHallucinationDetector()
    elif method == "hybrid":
        return HybridHallucinationDetector(use_vectara=False, use_nli=True)  # 禁用Vectara,使用NLI
    else:
        raise ValueError(f"未知的检测方法: {method}")