File size: 20,699 Bytes
af83196 | 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 | """
AdaEvolve context builder for SkyDiscover.
Extends DefaultContextBuilder with AdaEvolve-specific prompt sections:
- Evaluator feedback from parent artifacts
- Paradigm breakthrough guidance
- Sibling context (previous mutations of the same parent)
- Error retry context
These are assembled into a ``search_guidance`` string and injected into
AdaEvolve-specific templates via the ``{search_guidance}`` placeholder.
"""
import logging
import re
from pathlib import Path
from typing import Any, Dict, List, Optional, Union
from skydiscover.config import Config
from skydiscover.context_builder.default import DefaultContextBuilder
from skydiscover.context_builder.utils import TemplateManager, prog_attr
from skydiscover.search.base_database import Program
from skydiscover.utils.metrics import compute_proxy_score, get_score
logger = logging.getLogger(__name__)
class AdaEvolveContextBuilder(DefaultContextBuilder):
"""
Context builder for AdaEvolve's adaptive evolutionary search.
Adds a ``{search_guidance}`` section to the prompt containing:
- Evaluator diagnostic feedback (from parent's artifacts)
- Paradigm breakthrough guidance (when search is globally stagnating)
- Sibling context (previous mutations of the same parent)
- Error retry context (when retrying after a failed generation)
The controller passes raw data via the ``context`` dict:
- ``context["paradigm"]``: paradigm dict or None
- ``context["siblings"]``: list of Program objects
- ``context["error_context"]``: error string or None
Evaluator feedback is extracted from the parent program's artifacts.
"""
def __init__(self, config: Config):
super().__init__(config)
default_templates = str(Path(__file__).parent.parent / "default" / "templates")
adaevolve_templates = str(Path(__file__).parent / "templates")
self.template_manager = TemplateManager(
default_templates, adaevolve_templates, self.context_config.template_dir
)
def _db_config(self) -> Any:
return getattr(self.config.search, "database", None)
def _is_multiobjective_enabled(self) -> bool:
return bool(getattr(self._db_config(), "pareto_objectives", None) or [])
def _objective_descriptions(self) -> List[str]:
db_config = self._db_config()
higher_is_better = getattr(db_config, "higher_is_better", None) or {}
descriptions = []
for objective in getattr(db_config, "pareto_objectives", None) or []:
direction = "maximize" if higher_is_better.get(objective, True) else "minimize"
descriptions.append(f"{objective} ({direction})")
return descriptions
def _metric_to_maximization_value(self, metric_name: str, value: Any) -> Optional[float]:
from skydiscover.utils.metrics import normalize_metric_value
higher_is_better = getattr(self._db_config(), "higher_is_better", None) or {}
return normalize_metric_value(metric_name, value, higher_is_better)
_PROGRESS_SCORE_MISSING = float("-inf")
def _get_progress_score(self, metrics: Dict[str, Any]) -> float:
"""Scalar proxy used only for prompt-side progress descriptions.
Returns ``_PROGRESS_SCORE_MISSING`` (``-inf``) for empty/missing metrics
so that callers can distinguish "no data" from "score is zero".
"""
db_config = self._db_config()
pareto_objectives = getattr(db_config, "pareto_objectives", None) or None
return compute_proxy_score(
metrics,
fitness_key=getattr(db_config, "fitness_key", None),
pareto_objectives=pareto_objectives,
higher_is_better=getattr(db_config, "higher_is_better", None) or {},
)
def _task_objective_text(self) -> str:
subject = (
"prompt" if (self.config.language or "").lower() in ("text", "prompt") else "program"
)
if not self._is_multiobjective_enabled():
return f"Suggest improvements to the {subject} that will improve its COMBINED_SCORE."
return (
f"Suggest improvements to the {subject} that improve its Pareto trade-offs across: "
+ ", ".join(self._objective_descriptions())
+ "."
)
def _diversity_dimensions_text(self) -> str:
if not self._is_multiobjective_enabled():
return "The system maintains diversity across these dimensions: score, complexity."
return "The system maintains diversity across Pareto trade-offs, novelty, and solution structure."
def _diversity_note_text(self) -> str:
if not self._is_multiobjective_enabled():
return "Different solutions with similar combined_score but different features are valuable."
return "Different solutions with similar overall trade-offs but different objective balances are valuable."
def build_prompt(
self,
current_program: Union[Program, Dict[str, Program]],
context: Dict[str, Any] = None,
**kwargs: Any,
) -> Dict[str, str]:
"""
Build prompt with AdaEvolve-specific search guidance.
Computes the ``search_guidance`` string from AdaEvolve context keys,
then delegates to the parent's ``build_prompt`` which fills the
``{search_guidance}`` placeholder in AdaEvolve templates.
"""
context = context or {}
# Build the search guidance from AdaEvolve-specific context
search_guidance = self._build_search_guidance(current_program, context)
# Override any caller-supplied search_guidance with our computed one
kwargs.pop("search_guidance", None)
# Pass search_guidance through **kwargs to template.format()
result = super().build_prompt(
current_program,
context,
search_guidance=search_guidance,
task_objective=self._task_objective_text(),
diversity_dimensions=self._diversity_dimensions_text(),
diversity_note=self._diversity_note_text(),
**kwargs,
)
return result
# =========================================================================
# Suppress default artifact feedback rendering
# =========================================================================
def _format_current_program(
self,
current_program: Union[Program, Dict[str, Program]],
language: str,
) -> str:
"""Override to suppress artifacts["feedback"] from {current_program}.
AdaEvolve renders evaluator feedback explicitly via _build_search_guidance
into {search_guidance}, so we strip it here to avoid duplication.
"""
# Remove feedback from artifacts so parent renderer skips it (rendered via search_guidance instead)
if isinstance(current_program, dict):
program = list(current_program.values())[0]
else:
program = current_program
artifacts = getattr(program, "artifacts", None)
saved_feedback = None
if isinstance(artifacts, dict) and "feedback" in artifacts:
saved_feedback = artifacts.pop("feedback")
try:
return super()._format_current_program(current_program, language)
finally:
if saved_feedback is not None and isinstance(artifacts, dict):
artifacts["feedback"] = saved_feedback
# =========================================================================
# Search Guidance Assembly
# =========================================================================
def _build_search_guidance(
self,
current_program: Union[Program, Dict[str, Program]],
context: Dict[str, Any],
) -> str:
"""
Assemble all AdaEvolve-specific guidance sections into one string.
Sections are included in priority order:
1. Evaluator feedback (highest value — shows why parent fails)
2. Paradigm breakthrough guidance (when globally stagnating)
3. Sibling context (previous mutations of this parent)
4. Error retry context (when retrying after failure)
"""
# Extract parent program from current_program dict
if isinstance(current_program, dict):
parent_program = list(current_program.values())[0]
else:
parent_program = current_program
language = self.config.language or "python"
paradigm = context.get("paradigm")
siblings = context.get("siblings", [])
error_context = context.get("error_context")
sections: List[str] = []
# 1. Evaluator feedback from parent artifacts
feedback_section = self._format_evaluator_feedback(parent_program)
if feedback_section:
sections.append(feedback_section)
# 2. Paradigm breakthrough guidance
if paradigm:
sections.append(self._format_paradigm_guidance(paradigm, language))
# 3. Sibling context
if siblings:
sibling_section = self._format_sibling_context(siblings, parent_program)
if sibling_section:
sections.append(sibling_section)
# 4. Error retry context
if error_context:
sections.append(self._format_error_context(error_context))
if not sections:
return ""
return "\n\n".join(sections)
def _identify_improvement_areas(
self,
current_program: str,
metrics: Dict[str, float],
previous_programs: List[Program],
) -> str:
"""Generate improvement bullets for scalar or Pareto mode."""
if not self._is_multiobjective_enabled():
return super()._identify_improvement_areas(current_program, metrics, previous_programs)
improvement_areas = [
"Focus on Pareto trade-offs across: " + ", ".join(self._objective_descriptions())
]
current_score = self._get_progress_score(metrics)
if previous_programs:
prev_metrics = prog_attr(previous_programs[-1], "metrics", {}) or {}
prev_score = self._get_progress_score(prev_metrics)
# Only show delta text when both scores are valid (not missing).
missing = self._PROGRESS_SCORE_MISSING
if current_score != missing and prev_score != missing:
if current_score > prev_score + 1e-6:
improvement_areas.append(
f"Pareto proxy improved: {prev_score:.4f} -> {current_score:.4f}"
)
elif current_score < prev_score - 1e-6:
improvement_areas.append(
f"Pareto proxy declined: {prev_score:.4f} -> {current_score:.4f}. Revisit recent trade-offs."
)
else:
improvement_areas.append(f"Pareto proxy unchanged at {current_score:.4f}")
elif current_score != missing:
improvement_areas.append(f"Pareto proxy at {current_score:.4f} (first measurement)")
threshold = self.context_config.suggest_simplification_after_chars
if threshold and len(current_program) > threshold:
improvement_areas.append(
f"Consider simplifying - solution length exceeds {threshold} characters"
)
return "\n".join(f"- {area}" for area in improvement_areas)
# =========================================================================
# Section Formatters
# =========================================================================
@staticmethod
def _format_evaluator_feedback(parent_program: Program) -> Optional[str]:
"""
Format evaluator feedback from parent's artifacts.
The evaluator may return diagnostic feedback (e.g. analysis of failed
examples) in artifacts["feedback"]. This is injected into the prompt
so the LLM can make targeted improvements instead of guessing.
"""
artifacts = getattr(parent_program, "artifacts", None)
if not artifacts:
return None
feedback = artifacts.get("feedback")
if not feedback or not isinstance(feedback, str):
return None
# Truncate very long feedback to keep prompt focused
max_len = 2000
if len(feedback) > max_len:
feedback = feedback[:max_len] + "\n... (truncated)"
return (
"## EVALUATOR FEEDBACK ON CURRENT PROGRAM\n"
"The evaluator analyzed cases where the current program failed "
"and produced the following diagnostic feedback. "
"Use this to make targeted improvements:\n\n"
f"{feedback}"
)
@staticmethod
def _format_paradigm_guidance(paradigm: Dict[str, Any], language: str) -> str:
"""
Format paradigm breakthrough guidance for the LLM.
Uses different framing for prompt optimization vs code optimization.
"""
is_prompt_opt = (language or "").lower() in ("text", "prompt")
idea = paradigm.get("idea", "N/A")
description = paradigm.get("description", "N/A")
target = paradigm.get("what_to_optimize", "score")
cautions = paradigm.get("cautions", "N/A")
approach_type = paradigm.get("approach_type", "N/A")
if is_prompt_opt:
header = "## BREAKTHROUGH STRATEGY - APPLY THIS"
intro = "The search has stagnated globally. You MUST apply this breakthrough prompt strategy:"
fields = (
f"**STRATEGY:** {idea}\n\n"
f"**HOW TO APPLY:**\n{description}\n\n"
f"**TARGET:** {target}\n\n"
f"**CAUTIONS:** {cautions}\n\n"
f"**APPROACH TYPE:** {approach_type}"
)
critical_bullets = (
"- You MUST rewrite the prompt using this strategy\n"
"- The strategy must be reflected in the actual prompt structure and content\n"
"- Keep the prompt clear and well-structured\n"
"- Do not add unnecessary verbosity — every sentence should serve a purpose\n"
"- Ensure the prompt still addresses the core task"
)
else:
header = "## BREAKTHROUGH IDEA - IMPLEMENT THIS"
intro = "The search has stagnated globally. You MUST implement this breakthrough idea:"
fields = (
f"**IDEA:** {idea}\n\n"
f"**HOW TO IMPLEMENT:**\n{description}\n\n"
f"**TARGET METRIC:** {target}\n\n"
f"**CAUTIONS:** {cautions}\n\n"
f"**APPROACH TYPE:** {approach_type}"
)
critical_bullets = (
"- You MUST implement the breakthrough idea\n"
"- Ensure the paradigm is actually used in your implementation (not just mentioned in comments)\n"
"- Correctness is essential - your implementation must be correct and functional\n"
"- Verify output format matches evaluator requirements\n"
"- Make purposeful changes that implement the idea\n"
"- Test your implementation logic carefully"
)
return f"{header}\n\n{intro}\n\n{fields}\n\n**CRITICAL:**\n{critical_bullets}"
def _format_sibling_context(
self, siblings: List[Program], parent_program: Program
) -> Optional[str]:
"""
Format sibling context showing previous mutations of the parent.
Shows what mutations have been tried before, whether they improved
or regressed, so the LLM can avoid repeating failed approaches.
"""
if not siblings:
return None
parent_fitness = self._get_progress_score(getattr(parent_program, "metrics", {}))
missing = self._PROGRESS_SCORE_MISSING
improved, regressed, unchanged = 0, 0, 0
entries: List[str] = []
for i, child in enumerate(siblings, 1):
child_fitness = self._get_progress_score(getattr(child, "metrics", {}))
if parent_fitness == missing or child_fitness == missing:
entries.append(f" {i}. (metrics unavailable) [UNKNOWN]")
unchanged += 1
continue
delta = child_fitness - parent_fitness
if delta > 0.001:
status = "IMPROVED"
improved += 1
elif delta < -0.001:
status = "REGRESSED"
regressed += 1
else:
status = "NO CHANGE"
unchanged += 1
entries.append(
f" {i}. {parent_fitness:.4f} -> {child_fitness:.4f} " f"({delta:+.4f}) [{status}]"
)
lines = [
"## PREVIOUS ATTEMPTS ON THIS PARENT",
f"Summary: {improved} improved, {unchanged} unchanged, {regressed} regressed",
*entries,
"Avoid repeating approaches that didn't work.",
]
return "\n".join(lines)
def _format_previous_attempts(
self, previous_programs: List[Program], num_previous_attempts: int = 3
) -> str:
"""Format recent attempts using AdaEvolve's scalar proxy in Pareto mode."""
if not self._is_multiobjective_enabled():
return super()._format_previous_attempts(previous_programs, num_previous_attempts)
if not previous_programs:
return "No previous attempts yet."
try:
previous_attempt_template = self.template_manager.get_template("previous_attempt")
except (ValueError, KeyError):
previous_attempt_template = "### Attempt {attempt_number}\n- Changes: {changes}\n- Metrics: {performance}\n- Outcome: {outcome}"
previous_programs = sorted(
previous_programs,
key=lambda program: self._get_progress_score(prog_attr(program, "metrics", {}) or {}),
reverse=True,
)
selected = previous_programs[: min(num_previous_attempts, len(previous_programs))]
lines = []
for i, program in enumerate(reversed(selected)):
attempt_number = len(selected) - i
metadata = prog_attr(program, "metadata", {}) or {}
metrics = prog_attr(program, "metrics", {}) or {}
changes = metadata.get("changes", "Unknown changes")
performance_parts = []
for name, value in metrics.items():
if not isinstance(value, bool) and isinstance(value, (int, float)):
try:
performance_parts.append(f"{name}: {value:.4f}")
except (ValueError, TypeError):
performance_parts.append(f"{name}: {value}")
else:
performance_parts.append(f"{name}: {value}")
performance_str = ", ".join(performance_parts) if performance_parts else "No metrics"
parent_metrics = metadata.get("parent_metrics", {})
outcome = self._determine_outcome(metrics, parent_metrics)
lines.append(
previous_attempt_template.format(
attempt_number=attempt_number,
changes=changes,
performance=performance_str,
outcome=outcome,
)
+ "\n\n"
)
return "".join(lines)
def _determine_outcome(
self, program_metrics: Dict[str, Any], parent_metrics: Dict[str, Any]
) -> str:
"""Describe attempt outcome using the configured scalar proxy in Pareto mode."""
if not self._is_multiobjective_enabled():
return super()._determine_outcome(program_metrics, parent_metrics)
prog_value = self._get_progress_score(program_metrics)
parent_value = self._get_progress_score(parent_metrics)
missing = self._PROGRESS_SCORE_MISSING
if prog_value == missing or parent_value == missing:
return "Insufficient metrics for comparison"
if prog_value > parent_value + 1e-6:
return "Improvement in Pareto proxy"
if prog_value < parent_value - 1e-6:
return "Regression in Pareto proxy"
return "No change in Pareto proxy"
@staticmethod
def _format_error_context(error_context: str) -> str:
"""Format retry error context."""
return (
"## RETRY CONTEXT\n"
f"Previous attempt failed with error:\n```\n{error_context}\n```\n"
"Please fix this issue in your response."
)
|