Spaces:
Paused
Paused
File size: 9,528 Bytes
0d3f7cc | 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 | """Constraint gates for GEPA self-evolution pipeline.
Validates that evolved skills remain safe, semantically similar,
within size limits, and pass regression tests.
"""
from __future__ import annotations
import logging
import re
from typing import Any
logger = logging.getLogger(__name__)
class SizeGate:
"""Validates that evolved content stays within token size limits."""
def __init__(self, max_skill_tokens: int = 8192, max_prompt_tokens: int = 4096) -> None:
self.max_skill_tokens = max_skill_tokens
self.max_prompt_tokens = max_prompt_tokens
def check(self, content: str, content_type: str = "skill") -> dict[str, Any]:
"""Check if content is within size limits."""
estimated_tokens = len(content.split())
max_tokens = (
self.max_skill_tokens if content_type == "skill" else self.max_prompt_tokens
)
passed = estimated_tokens <= max_tokens
result: dict[str, Any] = {
"passed": passed,
"estimated_tokens": estimated_tokens,
"max_tokens": max_tokens,
"action": "ok" if passed else "truncate_and_warn",
}
if not passed:
result["message"] = (
f"Content exceeds {max_tokens} tokens ({estimated_tokens}). "
f"Will be truncated."
)
return result
class SemanticGate:
"""Validates semantic similarity between original and evolved content."""
def __init__(self, min_similarity: float = 0.85) -> None:
self.min_similarity = min_similarity
self._embedder = None
async def check(self, original: str, evolved: str) -> dict[str, Any]:
"""Check semantic similarity between original and evolved content."""
similarity = self._compute_similarity(original, evolved)
passed = similarity >= self.min_similarity
result: dict[str, Any] = {
"passed": passed,
"similarity": similarity,
"min_similarity": self.min_similarity,
"action": "ok" if passed else "reject",
}
if not passed:
result["message"] = (
f"Semantic similarity {similarity:.3f} below threshold {self.min_similarity}. "
f"Evolution rejected to prevent catastrophic forgetting."
)
return result
def _compute_similarity(self, text1: str, text2: str) -> float:
"""Compute semantic similarity using word overlap as fallback."""
try:
from sentence_transformers import SentenceTransformer
if self._embedder is None:
self._embedder = SentenceTransformer("BAAI/bge-small-en-v1.5")
emb1 = self._embedder.encode(text1)
emb2 = self._embedder.encode(text2)
import numpy as np
return float(np.dot(emb1, emb2) / (np.linalg.norm(emb1) * np.linalg.norm(emb2)))
except ImportError:
return self._token_overlap_similarity(text1, text2)
def _token_overlap_similarity(self, text1: str, text2: str) -> float:
"""Fallback: compute similarity based on token overlap."""
tokens1 = set(text1.lower().split())
tokens2 = set(text2.lower().split())
if not tokens1 or not tokens2:
return 0.0
intersection = tokens1 & tokens2
union = tokens1 | tokens2
return len(intersection) / len(union) if union else 0.0
class SecurityGate:
"""Validates that evolved content doesn't violate security rules."""
SECURITY_PATTERNS: list[tuple[str, str, str]] = [
(r"(?i)(password|secret|api[_-]?key|token|credential)\s*[=:]\s*\S+", "credential_leak", "Credentials or secrets in output"),
(r"(?i)(drop\s+table|truncate\s+table|delete\s+from|shutdown|format)", "destructive_operation", "Destructive DB operation"),
(r"(?i)(ssn|social.security|credit.card|pan[_-]?\d|aadhaar)", "pii_leak", "PII in output"),
(r"(?i)(bypass|disable)\s*(audit|log|security)", "audit_bypass", "Audit log bypass"),
]
def __init__(self) -> None:
self.enabled = True
def check(self, content: str) -> dict[str, Any]:
"""Check content for security violations."""
violations: list[dict[str, str]] = []
for pattern, violation_type, description in self.SECURITY_PATTERNS:
matches = re.findall(pattern, content)
if matches:
violations.append({
"type": violation_type,
"description": description,
"match_count": len(matches),
})
passed = len(violations) == 0
result: dict[str, Any] = {
"passed": passed,
"violations": violations,
"action": "ok" if passed else "reject_with_report",
}
if not passed:
result["message"] = (
f"Security violations found: {', '.join(v['type'] for v in violations)}. "
f"Evolution rejected."
)
return result
class RegressionGate:
"""Validates that evolved content passes regression tests."""
def __init__(self, min_pass_rate: float = 0.90) -> None:
self.min_pass_rate = min_pass_rate
async def check(
self, evolved_skill: str, test_cases: list[dict[str, Any]]
) -> dict[str, Any]:
"""Check if evolved skill passes regression tests."""
if not test_cases:
return {
"passed": True,
"pass_rate": 1.0,
"min_pass_rate": self.min_pass_rate,
"action": "ok",
"message": "No test cases to run.",
}
passed = 0
total = len(test_cases)
results: list[dict[str, Any]] = []
for case in test_cases:
try:
case_passed = self._evaluate_case(evolved_skill, case)
results.append({
"case_id": case.get("id", "unknown"),
"passed": case_passed,
})
if case_passed:
passed += 1
except Exception as e:
results.append({
"case_id": case.get("id", "unknown"),
"passed": False,
"error": str(e),
})
pass_rate = passed / total if total > 0 else 1.0
passed_ok = pass_rate >= self.min_pass_rate
result: dict[str, Any] = {
"passed": passed_ok,
"pass_rate": pass_rate,
"min_pass_rate": self.min_pass_rate,
"results": results,
"action": "ok" if passed_ok else "reject",
}
if not passed_ok:
result["message"] = (
f"Regression pass rate {pass_rate:.2f} below minimum {self.min_pass_rate}. "
f"Evolution rejected."
)
return result
def _evaluate_case(self, _skill: str, case: dict[str, Any]) -> bool:
"""Evaluate a single test case by running it against the evolved skill."""
task = case.get("input", {}).get("task", "")
expected = case.get("expected_output", {})
if not task:
return True
# Simulate evaluation: check if evolved content contains key terms from the task
skill_lower = _skill.lower()
task_lower = task.lower()
# Basic semantic check: task keywords should appear in skill or vice versa
task_words = set(task_lower.split())
skill_words = set(skill_lower.split())
overlap = task_words & skill_words
# If there's meaningful overlap or no expected output to compare against, pass
if overlap or not expected:
return True
# If expected output specifies required fields, check they exist
if isinstance(expected, dict):
for key in expected:
if key.lower() not in skill_lower and key.lower() not in task_lower:
return False
return True
class ConstraintGates:
"""Aggregates all constraint gates for the evolution pipeline."""
def __init__(
self,
max_skill_tokens: int = 8192,
min_semantic_similarity: float = 0.85,
min_regression_pass_rate: float = 0.90,
) -> None:
self.size_gate = SizeGate(max_skill_tokens=max_skill_tokens)
self.semantic_gate = SemanticGate(min_similarity=min_semantic_similarity)
self.security_gate = SecurityGate()
self.regression_gate = RegressionGate(min_pass_rate=min_regression_pass_rate)
async def check_all(
self,
original: str,
evolved: str,
content_type: str = "skill",
test_cases: list[dict[str, Any]] | None = None,
) -> dict[str, Any]:
"""Run all constraint gates."""
gates: dict[str, Any] = {}
gates["size"] = self.size_gate.check(evolved, content_type)
gates["semantic"] = await self.semantic_gate.check(original, evolved)
gates["security"] = self.security_gate.check(evolved)
if test_cases is not None:
gates["regression"] = await self.regression_gate.check(evolved, test_cases)
all_passed = all(g.get("passed", False) for g in gates.values())
return {
"passed": all_passed,
"gates": gates,
"failed_gates": [
name for name, g in gates.items() if not g.get("passed", False)
],
}
|