File size: 14,259 Bytes
32aefdf
 
 
 
 
 
 
 
 
 
 
 
714bdb1
32aefdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a6fc1b0
32aefdf
a6fc1b0
32aefdf
 
 
 
 
f1dccc0
32aefdf
f1dccc0
 
 
32aefdf
 
 
 
 
 
 
f1dccc0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
32aefdf
f1dccc0
32aefdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
RAG Evaluation Module
Comprehensive evaluation metrics for Retrieval-Augmented Generation systems.
"""

import json
import hashlib
from datetime import datetime
from typing import Optional, List, Dict, Any
from dataclasses import dataclass, asdict
import numpy as np
from pathlib import Path
import requests


@dataclass
class EvaluationResult:
    """Single evaluation result for a query-answer pair."""
    query: str
    answer: str
    source_docs: List[str]
    
    # Retrieval metrics
    num_retrieved: int
    retrieval_precision: float
    retrieval_recall: float
    rank_position: int  # Position of correct doc in ranked results
    
    # Generation metrics
    rouge_l: float  # Token-level overlap
    bert_score: float  # Semantic similarity
    answer_relevance: float  # Is answer relevant to query?
    
    # Faithfulness metrics
    faithfulness: float  # Is answer grounded in sources?
    hallucination_detected: bool
    source_attribution_score: float  # % of answer backed by sources
    
    # Performance metrics
    latency_ms: float
    tokens_used: int
    cost_cents: float
    
    # Metadata
    timestamp: str = ""
    eval_id: str = ""
    
    def __post_init__(self):
        if not self.timestamp:
            self.timestamp = datetime.now().isoformat()
        if not self.eval_id:
            # Generate unique ID from query hash
            self.eval_id = hashlib.md5(f"{self.query}{self.timestamp}".encode()).hexdigest()[:8]
    
    def to_dict(self) -> Dict[str, Any]:
        """Convert to dictionary."""
        data = asdict(self)
        # data['hallucination_detected'] = int(data['hallucination_detected'])
        return data


class RAGEvaluator:
    """Main evaluation engine for RAG systems."""
    
    def __init__(self, store_results: bool = True, results_dir: str = "evaluation_results"):
        """
        Initialize evaluator.
        
        Args:
            store_results: Whether to store results to disk
            results_dir: Directory to store evaluation results
        """
        BASE_DIR = Path(__file__).resolve().parents[2]
        self.store_results = store_results
        self.results_dir = BASE_DIR / results_dir
        self.results_dir.mkdir(exist_ok=True)
        self.results: List[EvaluationResult] = []
        self._load_existing_results()
    
    def _load_existing_results(self):
        """Load existing results from disk or HF raw URL."""
        results_file = self.results_dir / "results.jsonl"
        loaded = False

        # Try local file first
        if results_file.exists():
            try:
                with open(results_file, 'r') as f:
                    for line in f:
                        data = json.loads(line)
                        data['hallucination_detected'] = bool(data['hallucination_detected'])
                        self.results.append(EvaluationResult(**data))
                loaded = True
            except Exception as e:
                print(f"Warning: Could not load local results: {e}")

        # If no local file or failed, try HF raw URL
        if not loaded:
            hf_url = "https://huggingface.co/spaces/aankitdas/doc-intelligence-rag/raw/main/evaluation_results/results.jsonl"
            try:
                resp = requests.get(hf_url)
                resp.raise_for_status()
                for line in resp.text.strip().split("\n"):
                    data = json.loads(line)
                    data['hallucination_detected'] = bool(data['hallucination_detected'])
                    self.results.append(EvaluationResult(**data))
                print(f"Loaded {len(self.results)} results from HF raw URL")
            except Exception as e:
                print(f"Warning: Could not load results from HF: {e}")
    
    def add_result(self, result: EvaluationResult) -> None:
        """Add evaluation result."""
        self.results.append(result)
        if self.store_results:
            self._save_result(result)
    
    def _save_result(self, result: EvaluationResult) -> None:
        """Save single result to disk."""
        results_file = self.results_dir / "results.jsonl"
        try:
            with open(results_file, 'a') as f:
                f.write(json.dumps(result.to_dict()) + '\n')
        except Exception as e:
            print(f"Warning: Could not save result: {e}")
    
    def compute_aggregate_metrics(self) -> Dict[str, Any]:
        """Compute aggregate metrics across all results."""
        if not self.results:
            return self._empty_metrics()
        
        results_data = [r.to_dict() for r in self.results]
        
        # Convert to numeric arrays
        retrieval_precision = np.array([r['retrieval_precision'] for r in results_data])
        retrieval_recall = np.array([r['retrieval_recall'] for r in results_data])
        rouge_l = np.array([r['rouge_l'] for r in results_data])
        bert_score = np.array([r['bert_score'] for r in results_data])
        faithfulness = np.array([r['faithfulness'] for r in results_data])
        answer_relevance = np.array([r['answer_relevance'] for r in results_data])
        latency = np.array([r['latency_ms'] for r in results_data])
        costs = np.array([r['cost_cents'] for r in results_data])
        rank_pos = np.array([r['rank_position'] for r in results_data])
        hallucinations = np.array([r['hallucination_detected'] for r in results_data])
        source_attr = np.array([r['source_attribution_score'] for r in results_data])
        
        # Calculate MRR (Mean Reciprocal Rank)
        mrr = np.mean(1.0 / rank_pos)
        
        return {
            # Retrieval Metrics
            "retrieval_precision_mean": float(np.mean(retrieval_precision)),
            "retrieval_precision_std": float(np.std(retrieval_precision)),
            "retrieval_recall_mean": float(np.mean(retrieval_recall)),
            "retrieval_recall_std": float(np.std(retrieval_recall)),
            "mrr": float(mrr),
            
            # Generation Metrics
            "rouge_l_mean": float(np.mean(rouge_l)),
            "rouge_l_std": float(np.std(rouge_l)),
            "bert_score_mean": float(np.mean(bert_score)),
            "bert_score_std": float(np.std(bert_score)),
            "answer_relevance_mean": float(np.mean(answer_relevance)),
            "answer_relevance_std": float(np.std(answer_relevance)),
            
            # Faithfulness Metrics
            "faithfulness_mean": float(np.mean(faithfulness)),
            "faithfulness_std": float(np.std(faithfulness)),
            "hallucination_rate": float(np.sum(hallucinations) / len(hallucinations)),
            "source_attribution_mean": float(np.mean(source_attr)),
            "source_attribution_std": float(np.std(source_attr)),
            
            # Performance Metrics
            "latency_p50": float(np.percentile(latency, 50)),
            "latency_p95": float(np.percentile(latency, 95)),
            "latency_p99": float(np.percentile(latency, 99)),
            "latency_mean": float(np.mean(latency)),
            "latency_std": float(np.std(latency)),
            "cost_per_query": float(np.mean(costs)),
            "total_cost": float(np.sum(costs)),
            
            # Metadata
            "total_evaluations": len(self.results),
            "timestamp": datetime.now().isoformat(),
        }
    
    def get_results_timeseries(self) -> Dict[str, List[Any]]:
        """Get results as timeseries for visualization."""
        results_data = [r.to_dict() for r in self.results]
        
        if not results_data:
            return {}
        
        timeseries = {
            "query_idx": list(range(len(results_data))),
            "retrieval_precision": [r['retrieval_precision'] for r in results_data],
            "retrieval_recall": [r['retrieval_recall'] for r in results_data],
            "rouge_l": [r['rouge_l'] for r in results_data],
            "bert_score": [r['bert_score'] for r in results_data],
            "faithfulness": [r['faithfulness'] for r in results_data],
            "answer_relevance": [r['answer_relevance'] for r in results_data],
            "latency_ms": [r['latency_ms'] for r in results_data],
            "hallucination": [int(r['hallucination_detected']) for r in results_data],
        }
        
        return timeseries
    
    def get_failure_analysis(self) -> Dict[str, Any]:
        """Analyze failure modes."""
        if not self.results:
            return self._empty_failure_analysis()
        
        results_data = [r.to_dict() for r in self.results]
        
        # Define failure thresholds
        low_retrieval_threshold = np.median([r['retrieval_precision'] for r in results_data]) * 0.7
        low_generation_threshold = np.median([r['bert_score'] for r in results_data]) * 0.7
        low_faithfulness_threshold = 0.8
        
        failures = {
            "hallucinations": [],
            "low_retrieval": [],
            "low_generation": [],
            "low_faithfulness": [],
        }
        
        for r in results_data:
            if r['hallucination_detected']:
                failures["hallucinations"].append({
                    "eval_id": r['eval_id'],
                    "query": r['query'][:100],
                    "score": r['faithfulness']
                })
            
            if r['retrieval_precision'] < low_retrieval_threshold:
                failures["low_retrieval"].append({
                    "eval_id": r['eval_id'],
                    "query": r['query'][:100],
                    "score": r['retrieval_precision']
                })
            
            if r['bert_score'] < low_generation_threshold:
                failures["low_generation"].append({
                    "eval_id": r['eval_id'],
                    "query": r['query'][:100],
                    "score": r['bert_score']
                })
            
            if r['faithfulness'] < low_faithfulness_threshold:
                failures["low_faithfulness"].append({
                    "eval_id": r['eval_id'],
                    "query": r['query'][:100],
                    "score": r['faithfulness']
                })
        
        return {
            "total_failures": sum(len(v) for v in failures.values()),
            "failure_modes": {k: len(v) for k, v in failures.items()},
            "failure_details": failures,
        }
    
    def get_percentile_analysis(self) -> Dict[str, Any]:
        """Get percentile analysis for performance metrics."""
        if not self.results:
            return {}
        
        results_data = [r.to_dict() for r in self.results]
        
        metrics_to_analyze = {
            "retrieval_precision": [r['retrieval_precision'] for r in results_data],
            "retrieval_recall": [r['retrieval_recall'] for r in results_data],
            "rouge_l": [r['rouge_l'] for r in results_data],
            "bert_score": [r['bert_score'] for r in results_data],
            "faithfulness": [r['faithfulness'] for r in results_data],
            "latency_ms": [r['latency_ms'] for r in results_data],
        }
        
        percentile_analysis = {}
        for metric_name, values in metrics_to_analyze.items():
            percentile_analysis[metric_name] = {
                "p10": float(np.percentile(values, 10)),
                "p25": float(np.percentile(values, 25)),
                "p50": float(np.percentile(values, 50)),
                "p75": float(np.percentile(values, 75)),
                "p90": float(np.percentile(values, 90)),
                "p95": float(np.percentile(values, 95)),
                "p99": float(np.percentile(values, 99)),
            }
        
        return percentile_analysis
    
    def export_to_csv(self, filepath: str) -> None:
        """Export results to CSV."""
        if not self.results:
            print("No results to export")
            return
        
        import csv
        
        results_data = [r.to_dict() for r in self.results]
        
        if results_data:
            keys = results_data[0].keys()
            with open(filepath, 'w', newline='') as f:
                writer = csv.DictWriter(f, fieldnames=keys)
                writer.writeheader()
                writer.writerows(results_data)
            print(f"Exported {len(results_data)} results to {filepath}")
    
    def reset(self) -> None:
        """Clear all results."""
        self.results = []
        results_file = self.results_dir / "results.jsonl"
        if results_file.exists():
            results_file.unlink()
    
    @staticmethod
    def _empty_metrics() -> Dict[str, Any]:
        """Return empty metrics structure."""
        return {
            "retrieval_precision_mean": 0.0,
            "retrieval_precision_std": 0.0,
            "retrieval_recall_mean": 0.0,
            "retrieval_recall_std": 0.0,
            "mrr": 0.0,
            "rouge_l_mean": 0.0,
            "rouge_l_std": 0.0,
            "bert_score_mean": 0.0,
            "bert_score_std": 0.0,
            "answer_relevance_mean": 0.0,
            "answer_relevance_std": 0.0,
            "faithfulness_mean": 0.0,
            "faithfulness_std": 0.0,
            "hallucination_rate": 0.0,
            "source_attribution_mean": 0.0,
            "source_attribution_std": 0.0,
            "latency_p50": 0.0,
            "latency_p95": 0.0,
            "latency_p99": 0.0,
            "latency_mean": 0.0,
            "latency_std": 0.0,
            "cost_per_query": 0.0,
            "total_cost": 0.0,
            "total_evaluations": 0,
            "timestamp": datetime.now().isoformat(),
        }
    
    @staticmethod
    def _empty_failure_analysis() -> Dict[str, Any]:
        """Return empty failure analysis."""
        return {
            "total_failures": 0,
            "failure_modes": {
                "hallucinations": 0,
                "low_retrieval": 0,
                "low_generation": 0,
                "low_faithfulness": 0,
            },
            "failure_details": {
                "hallucinations": [],
                "low_retrieval": [],
                "low_generation": [],
                "low_faithfulness": [],
            },
        }