Spaces:
Running
Running
File size: 22,918 Bytes
7c3f0ce | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 | """
Forge Engine - Main orchestrator for the multi-agent reasoning forge.
Coordinates the full forge cycle:
concept -> problem_generator -> each agent analyzes -> critic evaluates
-> (feedback loop: weak agents revise) -> synthesis_engine -> training example
Supports three modes:
1. forge_single() — Original single-pass (fast, good for bulk generation)
2. forge_with_feedback() — Closed critic loop (agents revise based on scores)
3. forge_with_debate() — Multi-turn debate (agents challenge each other)
Outputs JSONL training data in OpenAI chat format.
"""
import json
import os
import sys
import random
from typing import TextIO
from reasoning_forge.agents.newton_agent import NewtonAgent
from reasoning_forge.agents.quantum_agent import QuantumAgent
from reasoning_forge.agents.ethics_agent import EthicsAgent
from reasoning_forge.agents.philosophy_agent import PhilosophyAgent
from reasoning_forge.agents.davinci_agent import DaVinciAgent
from reasoning_forge.agents.empathy_agent import EmpathyAgent
from reasoning_forge.agents.critic_agent import CriticAgent
from reasoning_forge.synthesis_engine import SynthesisEngine
from reasoning_forge.problem_generator import ProblemGenerator
from reasoning_forge.epistemic_metrics import EpistemicMetrics
SYSTEM_PROMPT = (
"You are Codette, a multi-perspective reasoning AI. You analyze concepts "
"by examining them through multiple intellectual lenses -- physics, "
"philosophy, ethics, creative invention, and human empathy -- then "
"synthesize a unified understanding that is richer than any single "
"perspective. You think carefully, acknowledge uncertainty, and connect "
"abstract reasoning to concrete human experience."
)
# Score below which an agent gets sent back for revision
_REVISION_THRESHOLD = 0.6
class ForgeEngine:
"""Main orchestrator for multi-agent reasoning data generation."""
def __init__(self):
# Initialize all reasoning agents
self.newton = NewtonAgent()
self.quantum = QuantumAgent()
self.ethics = EthicsAgent()
self.philosophy = PhilosophyAgent()
self.davinci = DaVinciAgent()
self.empathy = EmpathyAgent()
self.critic = CriticAgent()
self.analysis_agents = [
self.newton,
self.quantum,
self.ethics,
self.philosophy,
self.davinci,
self.empathy,
]
# Initialize supporting engines
self.synthesis = SynthesisEngine()
self.problem_generator = ProblemGenerator()
self.epistemic = EpistemicMetrics()
def forge_single(self, concept: str) -> dict:
"""Run full forge cycle on one concept (original single-pass mode).
The cycle:
1. Generate reasoning problems from the concept.
2. Each analysis agent produces its perspective.
3. The critic evaluates the ensemble.
4. The synthesis engine combines everything.
5. Package as a training example.
Args:
concept: The concept text to forge.
Returns:
Training example dict in OpenAI chat format.
"""
# Step 1: Generate reasoning problems
problems = self.problem_generator.generate_problems(concept)
# Step 2: Each agent analyzes the concept
analyses = {}
for agent in self.analysis_agents:
analyses[agent.name] = agent.analyze(concept)
# Step 3: Critic evaluates the ensemble
critique = self.critic.evaluate_ensemble(concept, analyses)
# Step 4: Synthesis engine combines everything
synthesized_response = self.synthesis.synthesize(
concept, analyses, critique
)
# Step 5: Build the user prompt
if problems and random.random() < 0.5:
problem_type, problem_text = random.choice(problems)
user_content = problem_text
else:
user_content = (
f"Analyze this concept from multiple perspectives:\n\n{concept}"
)
# Step 6: Compute RC+xi epistemic metrics
epistemic_report = self.epistemic.full_epistemic_report(
analyses, synthesized_response
)
# Step 7: Package as training example
training_example = {
"messages": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": user_content},
{"role": "assistant", "content": synthesized_response},
],
"metadata": {
"concept": concept,
"agent_scores": critique.get("agent_scores", {}),
"overall_quality": critique.get("overall_quality", 0.0),
"problems_generated": len(problems),
"problem_types": [p[0] for p in problems],
"redundancies_found": len(critique.get("redundancies", [])),
"missing_perspectives": len(
critique.get("missing_perspectives", [])
),
"epistemic_tension": epistemic_report.get("tension_magnitude", 0),
"ensemble_coherence": epistemic_report.get("ensemble_coherence", 0),
"perspective_coverage": epistemic_report.get("perspective_coverage", {}),
"tension_productivity": epistemic_report.get("tension_productivity", {}),
},
}
return training_example
# -- Closed Critic Feedback Loop (new) ---------------------------------
def forge_with_feedback(
self,
concept: str,
max_revisions: int = 2,
) -> dict:
"""Run forge with closed critic feedback loop.
After initial analysis, the critic scores each agent. Agents scoring
below the revision threshold are sent back with specific critique
for a second attempt. The best version (original or revised) is kept.
Args:
concept: The concept text to forge.
max_revisions: Maximum revision rounds per weak agent.
Returns:
Training example dict with revision metadata.
"""
problems = self.problem_generator.generate_problems(concept)
# Initial analysis pass
analyses = {}
for agent in self.analysis_agents:
analyses[agent.name] = agent.analyze(concept)
revision_counts = {agent.name: 0 for agent in self.analysis_agents}
for revision_round in range(max_revisions):
critique = self.critic.evaluate_ensemble(concept, analyses)
agent_scores = critique.get("agent_scores", {})
suggestions = critique.get("improvement_suggestions", [])
# Find agents below threshold
weak_agents = [
agent for agent in self.analysis_agents
if agent_scores.get(agent.name, {}).get("combined", 1.0) < _REVISION_THRESHOLD
]
if not weak_agents:
break # All agents above threshold — converged
for agent in weak_agents:
score = agent_scores.get(agent.name, {})
# Build revision directive from critic feedback
directive = self._build_revision_directive(
agent.name, score, suggestions, concept
)
# Agent re-analyzes with the directive prepended to concept
revised = agent.analyze(f"{directive}\n\n{concept}")
# Keep revision only if it scores better (evaluate in full ensemble context)
old_score = score.get("combined", 0)
test_analyses = dict(analyses)
test_analyses[agent.name] = revised
new_critique = self.critic.evaluate_ensemble(
concept, test_analyses
)
new_score = new_critique.get("agent_scores", {}).get(
agent.name, {}
).get("combined", 0)
if new_score > old_score:
analyses[agent.name] = revised
revision_counts[agent.name] += 1
# Final critique and synthesis
final_critique = self.critic.evaluate_ensemble(concept, analyses)
synthesized = self.synthesis.synthesize(concept, analyses, final_critique)
epistemic_report = self.epistemic.full_epistemic_report(analyses, synthesized)
if problems and random.random() < 0.5:
problem_type, problem_text = random.choice(problems)
user_content = problem_text
else:
user_content = f"Analyze this concept from multiple perspectives:\n\n{concept}"
return {
"messages": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": user_content},
{"role": "assistant", "content": synthesized},
],
"metadata": {
"concept": concept,
"agent_scores": final_critique.get("agent_scores", {}),
"overall_quality": final_critique.get("overall_quality", 0.0),
"problems_generated": len(problems),
"revision_counts": revision_counts,
"total_revisions": sum(revision_counts.values()),
"epistemic_tension": epistemic_report.get("tension_magnitude", 0),
"ensemble_coherence": epistemic_report.get("ensemble_coherence", 0),
"tension_productivity": epistemic_report.get("tension_productivity", {}),
"forge_mode": "feedback_loop",
},
}
# -- Multi-Turn Debate (new) -------------------------------------------
def forge_with_debate(
self,
concept: str,
debate_rounds: int = 2,
) -> dict:
"""Run forge with multi-turn agent debate.
Each round:
1. All agents produce their analysis.
2. Random pairs are formed for cross-perspective challenge.
3. Each agent in a pair sees the other's analysis and produces
a response that engages with it.
4. Epistemic tension is tracked per round.
5. After all rounds, synthesis incorporates debate history.
Args:
concept: The concept text to forge.
debate_rounds: Number of debate rounds.
Returns:
Training example with debate history and tension decay metrics.
"""
problems = self.problem_generator.generate_problems(concept)
# Round 0: initial analyses
analyses = {}
for agent in self.analysis_agents:
analyses[agent.name] = agent.analyze(concept)
round_analyses = [dict(analyses)] # snapshot for tension tracking
debate_log = []
for round_num in range(debate_rounds):
# Form random pairs (odd agent out debates the first agent)
agents_shuffled = list(self.analysis_agents)
random.shuffle(agents_shuffled)
pairs = []
for i in range(0, len(agents_shuffled) - 1, 2):
pairs.append((agents_shuffled[i], agents_shuffled[i + 1]))
if len(agents_shuffled) % 2 == 1:
pairs.append((agents_shuffled[-1], agents_shuffled[0]))
round_debates = []
for agent_a, agent_b in pairs:
# Agent A sees B's analysis and responds
challenge_prompt = (
f"Another perspective on '{concept}' argues:\n\n"
f"{analyses[agent_b.name]}\n\n"
f"Respond to this from your {agent_a.perspective} perspective. "
f"Where do you agree, disagree, or see complementary insights?"
)
response_a = agent_a.analyze(challenge_prompt)
# Agent B sees A's response
counter_prompt = (
f"A {agent_a.perspective} perspective responded to your analysis "
f"of '{concept}':\n\n{response_a}\n\n"
f"Integrate their insights with your own view."
)
response_b = agent_b.analyze(counter_prompt)
# Update analyses with debate-enriched versions
analyses[agent_a.name] = response_a
analyses[agent_b.name] = response_b
round_debates.append({
"pair": f"{agent_a.name}_vs_{agent_b.name}",
"challenge": response_a[:200],
"counter": response_b[:200],
})
debate_log.append({
"round": round_num + 1,
"debates": round_debates,
})
round_analyses.append(dict(analyses))
# Track tension decay across rounds
convergence = self.epistemic.score_debate_convergence(round_analyses)
# Final critique and synthesis
critique = self.critic.evaluate_ensemble(concept, analyses)
synthesized = self.synthesis.synthesize(concept, analyses, critique)
epistemic_report = self.epistemic.full_epistemic_report(analyses, synthesized)
if problems and random.random() < 0.5:
problem_type, problem_text = random.choice(problems)
user_content = problem_text
else:
user_content = f"Analyze this concept from multiple perspectives:\n\n{concept}"
return {
"messages": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": user_content},
{"role": "assistant", "content": synthesized},
],
"metadata": {
"concept": concept,
"agent_scores": critique.get("agent_scores", {}),
"overall_quality": critique.get("overall_quality", 0.0),
"problems_generated": len(problems),
"debate_rounds": debate_rounds,
"debate_log": debate_log,
"tension_decay": convergence,
"epistemic_tension": epistemic_report.get("tension_magnitude", 0),
"ensemble_coherence": epistemic_report.get("ensemble_coherence", 0),
"tension_productivity": epistemic_report.get("tension_productivity", {}),
"forge_mode": "debate",
},
}
# -- Helpers -----------------------------------------------------------
def _build_revision_directive(
self,
agent_name: str,
score: dict,
suggestions: list,
concept: str,
) -> str:
"""Build a revision directive for a weak agent."""
parts = [
f"[REVISION REQUESTED for {agent_name}]",
f"Your previous analysis scored {score.get('combined', 0):.2f}/1.00.",
]
if score.get("logical_clarity", 1) < 0.5:
parts.append(
"Improve logical clarity: use connectives (therefore, because, however), "
"avoid vague language, structure your argument explicitly."
)
if score.get("conceptual_accuracy", 1) < 0.5:
parts.append(
"Improve conceptual accuracy: engage directly with the specific concept, "
"use domain vocabulary, avoid generic placeholder framing."
)
if suggestions:
parts.append(f"Critic suggests: {suggestions[0]}")
parts.append("Reanalyze with these improvements:")
return " ".join(parts)
def forge_batch(
self, concept: str, variants: int = 3
) -> list[dict]:
"""Generate multiple training examples from one concept.
Uses different problem framings and agent template selections
to produce varied training data from the same concept.
Args:
concept: The concept text.
variants: Number of variants to generate.
Returns:
List of training example dicts.
"""
examples = []
for _ in range(variants):
example = self.forge_single(concept)
examples.append(example)
return examples
def forge_dataset(
self,
concepts: list[str],
output_path: str,
variants_per_concept: int = 1,
verbose: bool = False,
) -> dict:
"""Run forge on a list of concepts and write JSONL output.
Args:
concepts: List of concept strings.
output_path: Path to output JSONL file.
variants_per_concept: Number of training examples per concept.
verbose: Whether to print progress.
Returns:
Summary dict with counts and quality statistics.
"""
os.makedirs(os.path.dirname(os.path.abspath(output_path)), exist_ok=True)
total_examples = 0
total_quality = 0.0
quality_scores = []
with open(output_path, "w", encoding="utf-8") as f:
for i, concept in enumerate(concepts):
if verbose:
print(
f"[{i + 1}/{len(concepts)}] Forging: "
f"{concept[:60]}{'...' if len(concept) > 60 else ''}",
file=sys.stderr,
)
for variant in range(variants_per_concept):
example = self.forge_single(concept)
quality = example["metadata"]["overall_quality"]
# Write the messages (without metadata) for training
training_record = {"messages": example["messages"]}
f.write(json.dumps(training_record, ensure_ascii=False) + "\n")
total_examples += 1
total_quality += quality
quality_scores.append(quality)
summary = {
"total_examples": total_examples,
"total_concepts": len(concepts),
"variants_per_concept": variants_per_concept,
"output_path": output_path,
"avg_quality": round(total_quality / max(1, total_examples), 3),
"min_quality": round(min(quality_scores) if quality_scores else 0, 3),
"max_quality": round(max(quality_scores) if quality_scores else 0, 3),
}
if verbose:
print(f"\nForge complete: {summary}", file=sys.stderr)
return summary
def forge_from_dataset(
self,
input_jsonl: str,
output_path: str,
concept_field: str = "text",
variants_per_concept: int = 1,
verbose: bool = False,
) -> dict:
"""Read an existing JSONL dataset and run forge on each entry.
Expects each line to be a JSON object with a text field containing
the concept. Supports common field names: 'text', 'concept',
'content', 'input', 'question', 'prompt'.
Args:
input_jsonl: Path to input JSONL file.
output_path: Path to output JSONL file.
concept_field: Name of the field containing the concept text.
variants_per_concept: Number of training examples per concept.
verbose: Whether to print progress.
Returns:
Summary dict with counts and quality statistics.
"""
# Candidate field names to try
candidate_fields = [
concept_field, "text", "concept", "content",
"input", "question", "prompt",
]
concepts = []
with open(input_jsonl, "r", encoding="utf-8") as f:
for line_num, line in enumerate(f, 1):
line = line.strip()
if not line:
continue
try:
record = json.loads(line)
except json.JSONDecodeError:
if verbose:
print(
f"Warning: skipping malformed JSON on line {line_num}",
file=sys.stderr,
)
continue
# Try candidate fields in order
concept_text = None
if isinstance(record, dict):
for field in candidate_fields:
if field in record and isinstance(record[field], str):
concept_text = record[field].strip()
break
# Fallback: if record has 'messages', extract user content
if concept_text is None and "messages" in record:
for msg in record["messages"]:
if msg.get("role") == "user":
concept_text = msg["content"].strip()
break
elif isinstance(record, str):
concept_text = record.strip()
if concept_text:
concepts.append(concept_text)
if verbose:
print(
f"Loaded {len(concepts)} concepts from {input_jsonl}",
file=sys.stderr,
)
return self.forge_dataset(
concepts,
output_path,
variants_per_concept=variants_per_concept,
verbose=verbose,
)
def forge_single_detailed(self, concept: str) -> dict:
"""Run forge cycle and return all intermediate outputs.
Useful for debugging, inspection, and quality analysis.
Args:
concept: The concept text.
Returns:
Dict with all intermediate results:
{
"concept": str,
"problems": [(type, text), ...],
"analyses": {agent_name: analysis_text, ...},
"critique": {...},
"synthesis": str,
"training_example": {...},
}
"""
problems = self.problem_generator.generate_problems(concept)
analyses = {}
for agent in self.analysis_agents:
analyses[agent.name] = agent.analyze(concept)
critique = self.critic.evaluate_ensemble(concept, analyses)
synthesized = self.synthesis.synthesize(concept, analyses, critique)
user_content = (
f"Analyze this concept from multiple perspectives:\n\n{concept}"
)
training_example = {
"messages": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": user_content},
{"role": "assistant", "content": synthesized},
],
}
return {
"concept": concept,
"problems": problems,
"analyses": analyses,
"critique": critique,
"synthesis": synthesized,
"training_example": training_example,
}
|