Spaces:
Running
Running
File size: 10,013 Bytes
24f95f0 | 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 | """
Prompt evolution through A/B testing and versioning.
Creates prompt variants, tests them, and promotes better versions.
"""
import json
import logging
from datetime import datetime
from pathlib import Path
from typing import Dict, Any, Optional, List
import uuid
logger = logging.getLogger(__name__)
class PromptOptimizer:
"""Manages prompt evolution and A/B testing."""
def __init__(self, data_dir: str, model_fn):
self.data_dir = Path(data_dir) / "prompt_versions"
self.data_dir.mkdir(parents=True, exist_ok=True)
self.model_fn = model_fn
async def create_prompt_variant(self, prompt_name: str, current_prompt: str, goal: str) -> Dict[str, Any]:
"""
Create a new prompt variant using LLM.
Args:
prompt_name: Name of the prompt (e.g., "research", "verifier")
current_prompt: Current prompt text
goal: Optimization goal (e.g., "improve clarity", "reduce tokens")
Returns:
Prompt variant with metadata
"""
logger.info(f"Creating prompt variant for {prompt_name} with goal: {goal}")
optimization_prompt = f"""You are a prompt engineering expert. Improve the following prompt based on this goal: {goal}
Current prompt:
{current_prompt}
Create an improved version that:
1. Maintains the same core functionality
2. {goal}
3. Is clear and actionable
Improved prompt:"""
try:
improved_prompt = await self.model_fn(optimization_prompt, max_tokens=2000)
variant = {
"id": str(uuid.uuid4()),
"prompt_name": prompt_name,
"version": self._get_next_version(prompt_name),
"prompt_text": improved_prompt,
"goal": goal,
"status": "testing",
"created_at": datetime.utcnow().isoformat(),
"test_count": 0,
"win_count": 0,
"win_rate": 0.0,
}
self._save_variant(variant)
logger.info(f"Created prompt variant: {variant['id']} (v{variant['version']})")
return variant
except Exception as e:
logger.error(f"Failed to create prompt variant: {e}")
raise
async def test_prompt_variant(self, variant_id: str, test_input: str, expected_quality: str) -> Dict[str, Any]:
"""
Test a prompt variant with sample input.
Args:
variant_id: Variant ID
test_input: Test input
expected_quality: Expected quality criteria
Returns:
Test results
"""
variant = self._load_variant(variant_id)
if not variant:
raise ValueError(f"Variant not found: {variant_id}")
logger.info(f"Testing prompt variant: {variant_id}")
try:
# Run the prompt with test input
test_prompt = variant["prompt_text"].replace("{input}", test_input)
output = await self.model_fn(test_prompt, max_tokens=1000)
# Evaluate quality
quality_score = await self._evaluate_quality(output, expected_quality)
# Update variant stats
variant["test_count"] += 1
if quality_score >= 0.7: # Consider it a "win" if quality >= 70%
variant["win_count"] += 1
variant["win_rate"] = variant["win_count"] / variant["test_count"]
self._save_variant(variant)
return {
"variant_id": variant_id,
"quality_score": quality_score,
"output": output,
"win_rate": variant["win_rate"],
}
except Exception as e:
logger.error(f"Failed to test prompt variant: {e}")
raise
def compare_prompts(self, variant_id_a: str, variant_id_b: str) -> Dict[str, Any]:
"""
Compare two prompt variants.
Args:
variant_id_a: First variant ID
variant_id_b: Second variant ID
Returns:
Comparison results
"""
variant_a = self._load_variant(variant_id_a)
variant_b = self._load_variant(variant_id_b)
if not variant_a or not variant_b:
raise ValueError("One or both variants not found")
return {
"variant_a": {
"id": variant_a["id"],
"version": variant_a["version"],
"win_rate": variant_a["win_rate"],
"test_count": variant_a["test_count"],
},
"variant_b": {
"id": variant_b["id"],
"version": variant_b["version"],
"win_rate": variant_b["win_rate"],
"test_count": variant_b["test_count"],
},
"winner": variant_a["id"] if variant_a["win_rate"] > variant_b["win_rate"] else variant_b["id"],
}
def promote_prompt(self, variant_id: str, min_tests: int = 10, min_win_rate: float = 0.7) -> bool:
"""
Promote a prompt variant to production.
Args:
variant_id: Variant ID
min_tests: Minimum number of tests required
min_win_rate: Minimum win rate required
Returns:
True if promoted, False otherwise
"""
variant = self._load_variant(variant_id)
if not variant:
raise ValueError(f"Variant not found: {variant_id}")
# Validate promotion criteria
if variant["test_count"] < min_tests:
logger.warning(f"Variant {variant_id} has insufficient tests: {variant['test_count']} < {min_tests}")
return False
if variant["win_rate"] < min_win_rate:
logger.warning(f"Variant {variant_id} has insufficient win rate: {variant['win_rate']} < {min_win_rate}")
return False
# Archive current production version
current_production = self._get_production_variant(variant["prompt_name"])
if current_production:
current_production["status"] = "archived"
current_production["archived_at"] = datetime.utcnow().isoformat()
self._save_variant(current_production)
# Promote variant
variant["status"] = "production"
variant["promoted_at"] = datetime.utcnow().isoformat()
self._save_variant(variant)
logger.info(f"Promoted prompt variant {variant_id} to production")
return True
def archive_prompt(self, variant_id: str):
"""Archive a prompt variant."""
variant = self._load_variant(variant_id)
if not variant:
raise ValueError(f"Variant not found: {variant_id}")
variant["status"] = "archived"
variant["archived_at"] = datetime.utcnow().isoformat()
self._save_variant(variant)
logger.info(f"Archived prompt variant: {variant_id}")
def list_versions(self, prompt_name: str) -> List[Dict[str, Any]]:
"""List all versions of a prompt."""
versions = []
for file_path in self.data_dir.glob(f"{prompt_name}_*.json"):
try:
with open(file_path, 'r') as f:
variant = json.load(f)
versions.append(variant)
except Exception as e:
logger.error(f"Failed to read variant {file_path}: {e}")
# Sort by version descending
versions.sort(key=lambda x: x.get("version", 0), reverse=True)
return versions
def _save_variant(self, variant: Dict[str, Any]):
"""Save a prompt variant to disk."""
file_path = self.data_dir / f"{variant['prompt_name']}_{variant['id']}.json"
with open(file_path, 'w') as f:
json.dump(variant, f, indent=2)
def _load_variant(self, variant_id: str) -> Optional[Dict[str, Any]]:
"""Load a prompt variant from disk."""
for file_path in self.data_dir.glob(f"*_{variant_id}.json"):
with open(file_path, 'r') as f:
return json.load(f)
return None
def _get_production_variant(self, prompt_name: str) -> Optional[Dict[str, Any]]:
"""Get the current production variant for a prompt."""
for file_path in self.data_dir.glob(f"{prompt_name}_*.json"):
try:
with open(file_path, 'r') as f:
variant = json.load(f)
if variant.get("status") == "production":
return variant
except Exception as e:
logger.error(f"Failed to read variant {file_path}: {e}")
return None
def _get_next_version(self, prompt_name: str) -> int:
"""Get the next version number for a prompt."""
versions = self.list_versions(prompt_name)
if not versions:
return 1
return max(v.get("version", 0) for v in versions) + 1
async def _evaluate_quality(self, output: str, expected_quality: str) -> float:
"""Evaluate output quality using LLM."""
eval_prompt = f"""Evaluate the quality of this output based on the criteria: {expected_quality}
Output:
{output}
Rate the quality from 0.0 to 1.0, where:
- 0.0 = Completely fails criteria
- 0.5 = Partially meets criteria
- 1.0 = Fully meets criteria
Provide only the numeric score:"""
try:
score_text = await self.model_fn(eval_prompt, max_tokens=10)
score = float(score_text.strip())
return max(0.0, min(1.0, score)) # Clamp to [0, 1]
except Exception as e:
logger.error(f"Failed to evaluate quality: {e}")
return 0.5 # Default to neutral score
|