File size: 13,784 Bytes
449c371
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import re
import random
from typing import List, Dict, Tuple, Optional
from dataclasses import dataclass
from openai import OpenAI

@dataclass
class EvaluationResult:
    """Structured result from regex evaluation."""
    score: float
    true_positives: List[Dict]
    false_positives: List[Dict]
    false_negatives: List[Dict]
    error_message: Optional[str] = None

class RegexOptimizerAgent:
    """
    An AI agent that iteratively optimizes regex patterns using LLM feedback.
    
    The agent uses a reflexion-based approach:
    1. Evaluates current regex on ground truth data
    2. Generates improvement suggestions via LLM
    3. Tests new regex and compares performance
    4. Learns from failures through critique
    """
    
    def __init__(self, ground_truth_data: List[Dict], openai_client: OpenAI, 
                 model: str = "gpt-4", temperature: float = 0.7):
        """
        Initialize the optimizer agent.
        
        Args:
            ground_truth_data: List of dicts with 'text' and 'expected' keys
            openai_client: OpenAI client instance
            model: Model name (e.g., 'gpt-4', 'gpt-3.5-turbo')
            temperature: Sampling temperature for generation
        """
        self.data = ground_truth_data
        self.client = openai_client
        self.model = model
        self.temperature = temperature
        self.best_regex = ""
        self.best_score = -1.0
        self.iteration_history = []

    def _evaluate_and_sample(self, regex_str: str, k: int = 3) -> EvaluationResult:
        """
        Evaluate regex performance and sample results from each category.
        
        Args:
            regex_str: Regular expression pattern to test
            k: Number of samples to return per category
            
        Returns:
            EvaluationResult with score and categorized samples
        """
        tp_samples, fp_samples, fn_samples = [], [], []
        correct_count = 0
        error_msg = None
        
        # Validate regex syntax
        try:
            compiled = re.compile(regex_str)
        except Exception as e:
            return EvaluationResult(0.0, [], [], [], str(e))

        # Test on all data points
        for item in self.data:
            raw_text = item['text']
            expected = str(item['expected']) if item['expected'] is not None else None
            
            match = compiled.search(raw_text)
            actual = match.group(0) if match else None
            
            is_correct = (actual == expected)
            
            result_payload = {
                "raw": raw_text,
                "truth": expected if expected else "NO_MATCH",
                "extracted": actual if actual else "NONE"
            }

            if is_correct:
                correct_count += 1
                tp_samples.append(result_payload)
            else:
                if actual is None and expected is not None:
                    fn_samples.append(result_payload)  # Missed match
                else:
                    fp_samples.append(result_payload)  # Wrong match

        score = correct_count / len(self.data) if self.data else 0
        
        # Sample to keep prompt manageable
        random.shuffle(tp_samples)
        random.shuffle(fp_samples)
        random.shuffle(fn_samples)

        return EvaluationResult(
            score=score,
            true_positives=tp_samples[:k],
            false_positives=fp_samples[:k],
            false_negatives=fn_samples[:k],
            error_message=error_msg
        )

    def _generate_failure_critique(self, attempted_regex: str, best_regex: str, 
                                   new_fp: List[Dict], new_fn: List[Dict], 
                                   error_msg: Optional[str]) -> str:
        """Generate critique explaining why attempted regex failed."""
        if error_msg:
            return f"SYNTAX ERROR: {error_msg}"

        prompt = f"""You are an expert regex analyst. Analyze why this regex attempt failed.

BEST PERFORMING REGEX (baseline): {best_regex}
FAILED ATTEMPT: {attempted_regex}

NEW ERRORS INTRODUCED:
False Positives (incorrectly matched): {new_fp[:2]}
False Negatives (failed to match): {new_fn[:2]}

Task: In 1-2 concise sentences, explain the core flaw in the attempted regex that caused it to perform worse than the baseline. Focus on what pattern logic was incorrect.

Critique:"""

        try:
            response = self.client.chat.completions.create(
                model=self.model,
                messages=[
                    {"role": "system", "content": "You are an expert regular expression analyst."},
                    {"role": "user", "content": prompt}
                ],
                temperature=self.temperature,
                max_tokens=150
            )
            return response.choices[0].message.content.strip()
        except Exception as e:
            return f"Failed to generate critique: {str(e)}"

    def _build_optimization_prompt(self, regex: str, score: float, 
                                   tp: List[Dict], fp: List[Dict], fn: List[Dict],
                                   failure_context: Optional[str] = None) -> str:
        """
        Build comprehensive prompt for regex improvement.
        
        Uses chain-of-thought reasoning and specific examples to guide the LLM.
        """
        def format_examples(examples: List[Dict], limit: int = 5) -> str:
            if not examples:
                return "None"
            lines = []
            for ex in examples[:limit]:
                lines.append(f"  β€’ Input: '{ex['raw']}'")
                lines.append(f"    Expected: {ex['truth']} | Got: {ex['extracted']}")
            return "\n".join(lines)

        prompt = f"""You are an expert Regular Expression engineer. Your task is to improve a regex pattern to achieve 100% accuracy on the test cases below.

CURRENT REGEX PATTERN: `{regex}`
CURRENT ACCURACY: {score:.1%} ({int(score * len(self.data))}/{len(self.data)} correct)

═══════════════════════════════════════════════════════════════

[βœ“ CORRECT MATCHES - These MUST continue working]
{format_examples(tp)}

[βœ— FALSE POSITIVES - Incorrectly matched (should return NO_MATCH)]
{format_examples(fp)}

[βœ— FALSE NEGATIVES - Failed to extract (should have matched)]
{format_examples(fn)}

═══════════════════════════════════════════════════════════════
"""

        if failure_context:
            prompt += f"""
⚠️  CRITICAL FEEDBACK FROM PREVIOUS ATTEMPT:
{failure_context}

Learn from this mistake and avoid repeating it.

═══════════════════════════════════════════════════════════════
"""

        prompt += """
INSTRUCTIONS:
1. Analyze the patterns in false positives and false negatives
2. Identify what distinguishes valid matches from invalid ones
3. Design a regex that captures ONLY the expected patterns
4. Ensure all current correct matches continue working

Think step-by-step:
- What do all the expected matches have in common?
- What makes the false positives different from true matches?
- What patterns are the false negatives missing?

Return ONLY the improved regex pattern inside backticks, like: `your_regex_here`
Do not include explanations, just the regex pattern.

Improved regex:"""

        return prompt

    def _extract_regex_from_response(self, text: str) -> str:
        """Extract regex pattern from LLM response."""
        # Try to find pattern in backticks
        match = re.search(r"`([^`]+)`", text)
        if match:
            return match.group(1)
        
        # Fallback: take first line if no backticks found
        lines = [line.strip() for line in text.split('\n') if line.strip()]
        return lines[0] if lines else text.strip()

    def run_optimization(self, initial_regex: str, max_iterations: int = 10, 
                        patience: int = 3, sample_size: int = 3) -> str:
        """
        Run the iterative optimization loop.
        
        Args:
            initial_regex: Starting regex pattern
            max_iterations: Maximum optimization iterations
            patience: Stop after N iterations without improvement
            sample_size: Number of examples to show LLM per category
            
        Returns:
            Best regex pattern found
        """
        current_regex = initial_regex
        result = self._evaluate_and_sample(current_regex, k=sample_size)
        
        self.best_regex = current_regex
        self.best_score = result.score
        no_improvement_counter = 0
        last_failure_critique = None

        print("=" * 60)
        print("REGEX OPTIMIZATION AGENT STARTED")
        print("=" * 60)
        print(f"Initial Regex: {initial_regex}")
        print(f"Baseline Accuracy: {result.score:.1%} ({int(result.score * len(self.data))}/{len(self.data)})")
        print()

        for iteration in range(max_iterations):
            # Early stopping conditions
            if self.best_score == 1.0:
                print("βœ“ PERFECT SCORE ACHIEVED!")
                break
            
            if no_improvement_counter >= patience:
                print(f"βŠ— Early stopping: No improvement for {patience} iterations")
                break

            print(f"{'─' * 60}")
            print(f"Iteration {iteration + 1}/{max_iterations}")
            print(f"{'─' * 60}")
            
            # Generate candidate regex
            prompt = self._build_optimization_prompt(
                current_regex, self.best_score,
                result.true_positives, result.false_positives, result.false_negatives,
                last_failure_critique
            )
            print(f"Candidate: {candidate}")
            
            # Evaluate candidate
            new_result = self._evaluate_and_sample(candidate, k=sample_size)
            new_score = new_result.score
            
            # Compare performance
            if new_score > self.best_score:
                improvement = new_score - self.best_score
                print(f"βœ“ IMPROVEMENT: {self.best_score:.1%} β†’ {new_score:.1%} (+{improvement:.1%})")
                
                self.best_score = new_score
                self.best_regex = candidate
                current_regex = candidate
                result = new_result
                no_improvement_counter = 0
                last_failure_critique = None
                
                self.iteration_history.append({
                    "iteration": iteration + 1,
                    "regex": candidate,
                    "score": new_score,
                    "improved": True
                })
            else:
                print(f"βœ— No improvement: {new_score:.1%} (best: {self.best_score:.1%})")
                no_improvement_counter += 1
                
                # Generate critique for next iteration
                last_failure_critique = self._generate_failure_critique(
                    candidate, self.best_regex,
                    new_result.false_positives, new_result.false_negatives,
                    new_result.error_message
                )
                print(f"Critique: {last_failure_critique}")
                
                # Revert to best regex
                current_regex = self.best_regex
                
                self.iteration_history.append({
                    "iteration": iteration + 1,
                    "regex": candidate,
                    "score": new_score,
                    "improved": False
                })
            
            print()

        return self.best_regex

    def get_final_report(self) -> str:
        """Generate summary report of optimization process."""
        report = f"""
{'=' * 60}
OPTIMIZATION COMPLETE
{'=' * 60}
Final Regex: {self.best_regex}
Final Accuracy: {self.best_score:.1%} ({int(self.best_score * len(self.data))}/{len(self.data)})
Total Iterations: {len(self.iteration_history)}
Improvements: {sum(1 for h in self.iteration_history if h['improved'])}
{'=' * 60}
"""
        return report


def main():
    """Example usage with numerical ID extraction."""
    
    # Example dataset: Extract order/reference IDs (numbers that aren't prices or emergency codes)
    training_data = [
        {"text": "Order 5521 was shipped", "expected": "5521"},
        {"text": "Ref ID: 990", "expected": "990"},
        {"text": "Customer 12", "expected": "12"},
        {"text": "Invoice #7834 pending", "expected": "7834"},
        {"text": "Total is $50.00", "expected": None},  # Should NOT match money
        {"text": "Call 911 for help", "expected": None},  # Should NOT match emergency
        {"text": "Price: $19.99", "expected": None},
        {"text": "Tracking: 88234", "expected": "88234"}
    ]

    # Initialize OpenAI client
    # Make sure to set your OPENAI_API_KEY environment variable
    client = OpenAI()  # Reads from OPENAI_API_KEY env variable
    
    # Initialize agent
    agent = RegexOptimizerAgent(
        training_data, 
        client,
        model="gpt-4",  # or "gpt-3.5-turbo" for faster/cheaper runs
        temperature=0.7
    )
    
    # Run optimization
    final_regex = agent.run_optimization(
        initial_regex=r"\d",
        max_iterations=10,
        patience=3,
        sample_size=3
    )
    
    # Print results
    print(agent.get_final_report())


if __name__ == "__main__":
    main()