File size: 20,582 Bytes
6835659 | 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 | from __future__ import annotations
# Phase-3C: Targeted Modality Retry Controller
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple
from src.coherence.drift_detector import detect_drift
from src.coherence.msci import compute_msci_v0
from src.coherence.reporting import build_final_assessment
from src.coherence.scorer import CoherenceScorer
from src.coherence.controller import route_retry
from src.coherence.retry.retry_si_a import retry_si_a
from src.coherence.retry.retry_st_i import retry_st_i
from src.embeddings.aligned_embeddings import AlignedEmbedder
from src.generators.audio.generator import AudioGenerator
from src.generators.image.generator import ImageRetrievalGenerator
from src.generators.text.generator import TextGenerator
from src.narrative.generator import NarrativeGenerator
from src.orchestrator.regeneration_policy import decide_regeneration
from src.orchestrator.run_manager import create_run_paths
from src.planner.council import SemanticPlanningCouncil
from src.planner.schema import SemanticPlan
from src.planner.schema_to_text import plan_to_canonical_text
from src.storage.metadata import write_run_metadata
@dataclass(frozen=True)
class RunOutput:
run_id: str
semantic_plan: Dict[str, Any]
merge_report: Dict[str, Any]
planner_outputs: Dict[str, Any]
narrative_structured: Dict[str, Any]
narrative_text: str
image_path: str
audio_path: str
scores: Dict[str, Any]
coherence: Dict[str, Any]
final_assessment: Dict[str, Any]
drift: Dict[str, bool]
attempts: int
decisions: List[Dict[str, Any]]
class Orchestrator:
def __init__(
self,
council: SemanticPlanningCouncil,
text_gen: TextGenerator,
image_gen: ImageRetrievalGenerator,
audio_gen: AudioGenerator,
msci_threshold: float = 0.42,
max_attempts: int = 4,
runs_dir: str = "runs",
):
self.council = council
self.text_gen = text_gen
self.image_gen = image_gen
self.audio_gen = audio_gen
self.msci_threshold = msci_threshold
self.max_attempts = max_attempts
self.runs_dir = runs_dir
self.embedder = AlignedEmbedder(target_dim=512)
self.narrative_generator = NarrativeGenerator()
self.coherence_scorer = CoherenceScorer()
def run(self, user_prompt: str) -> RunOutput:
paths = create_run_paths(self.runs_dir)
council_result = self.council.run(user_prompt)
if isinstance(council_result, SemanticPlan):
plan = council_result
merge_report = {
"agreement_score": 1.0,
"per_section_agreement": {},
"conflicts": {},
"notes": "unified_planner",
}
planner_outputs = {"unified": plan.model_dump()}
else:
plan = council_result.merged_plan
merge_report = {
"agreement_score": council_result.merge_report.agreement_score,
"per_section_agreement": council_result.merge_report.per_section_agreement,
"conflicts": council_result.merge_report.conflicts,
"notes": council_result.merge_report.notes,
}
planner_outputs = {
"plan_a": council_result.plan_a.model_dump(),
"plan_b": council_result.plan_b.model_dump(),
"plan_c": council_result.plan_c.model_dump(),
}
plan_text = plan_to_canonical_text(plan)
plan_embedding = self.embedder.embed_text(plan_text)
img_pool = self.image_gen.retrieve_top_k(plan_text, k=8)
if not img_pool:
index_path = getattr(self.image_gen, "index_path", None)
hint = f" Expected index at {index_path}." if index_path else ""
raise RuntimeError(
"No image candidates retrieved. Build the image index or switch to a"
f" generative image backend.{hint}"
)
best_state: Optional[
Tuple[float, str, str, str, Dict[str, Any], Dict[str, bool], int]
] = None
decisions: List[Dict[str, Any]] = []
retry_outcomes: List[Dict[str, Any]] = []
narrative_structured = self.narrative_generator.generate(plan.model_dump())
narrative = narrative_structured.combined_scene
image_path = img_pool[0][0]
audio_path = str(paths.audio_dir / "audio_attempt1.wav")
audio_prompt = (
f"{plan.scene_summary}. Soundscape: {', '.join(plan.audio_elements)}. "
f"Mood: {', '.join(plan.mood_emotion)}."
)
retry_analysis: List[Dict[str, Any]] = []
epsilon = 0.01
for attempt in range(1, self.max_attempts + 1):
if attempt == 1:
audio_result = self.audio_gen.generate(audio_prompt, audio_path)
audio_path = audio_result.audio_path
audio_backend = audio_result.backend
else:
last_scores = decisions[-1]["scores"]
last_coherence = decisions[-1].get("coherence", {})
classification = last_coherence.get("classification", {})
context = {
"semantic_plan": plan.model_dump(),
"narrative_structured": narrative_structured.model_dump(),
"plan_text": plan_text,
"image_path": image_path,
"audio_path": audio_path,
"image_generator": self.image_gen,
"audio_generator": self.audio_gen,
}
retry_action = None
retry_strategy = None
retry_metric = None
retry_trigger = classification.get("label")
handled_regen = False
if (
classification.get("label") == "MODALITY_FAILURE"
and classification.get("weakest_metric") == "st_i"
):
context = retry_st_i(context)
image_path = context.get("image") or context.get("image_path") or image_path
retry_strategy = "ALIGN_IMAGE_TO_TEXT"
retry_metric = "st_i"
retry_action = {
"regenerate": "image",
"failed_metric": "st_i",
"strategy": retry_strategy,
}
handled_regen = True
elif (
classification.get("label") == "MODALITY_FAILURE"
and classification.get("weakest_metric") == "si_a"
):
audio_retry_path = str(paths.audio_dir / f"audio_attempt{attempt}.wav")
context["audio_path"] = audio_retry_path
context = retry_si_a(context)
audio_path = context.get("audio") or context.get("audio_path") or audio_path
audio_backend = context.get("audio_backend")
retry_meta = context.get("retry", {})
retry_strategy = retry_meta.get("strategy", "ALIGN_AUDIO_TO_IMAGE")
retry_metric = "si_a"
retry_action = {
"regenerate": "audio",
"failed_metric": "si_a",
"strategy": retry_strategy,
}
handled_regen = True
else:
retry_action = route_retry(classification, context)
if retry_action and retry_action.get("regenerate") == "full":
retry_strategy = retry_action.get("strategy")
retry_metric = retry_action.get("failed_metric")
handled_regen = True
council_result = self.council.run(user_prompt)
if isinstance(council_result, SemanticPlan):
plan = council_result
merge_report = {
"agreement_score": 1.0,
"per_section_agreement": {},
"conflicts": {},
"notes": "unified_planner",
}
planner_outputs = {"unified": plan.model_dump()}
else:
plan = council_result.merged_plan
merge_report = {
"agreement_score": council_result.merge_report.agreement_score,
"per_section_agreement": council_result.merge_report.per_section_agreement,
"conflicts": council_result.merge_report.conflicts,
"notes": council_result.merge_report.notes,
}
planner_outputs = {
"plan_a": council_result.plan_a.model_dump(),
"plan_b": council_result.plan_b.model_dump(),
"plan_c": council_result.plan_c.model_dump(),
}
plan_text = plan_to_canonical_text(plan)
plan_embedding = self.embedder.embed_text(plan_text)
narrative_structured = self.narrative_generator.generate(plan.model_dump())
narrative = narrative_structured.combined_scene
img_pool = self.image_gen.retrieve_top_k(plan_text, k=8)
if not img_pool:
index_path = getattr(self.image_gen, "index_path", None)
hint = f" Expected index at {index_path}." if index_path else ""
raise RuntimeError(
"No image candidates retrieved. Build the image index or switch to a"
f" generative image backend.{hint}"
)
image_path = img_pool[0][0]
audio_prompt = (
f"{plan.scene_summary}. Soundscape: {', '.join(plan.audio_elements)}. "
f"Mood: {', '.join(plan.mood_emotion)}."
)
audio_path = str(paths.audio_dir / f"audio_attempt{attempt}.wav")
audio_result = self.audio_gen.generate(audio_prompt, audio_path)
audio_path = audio_result.audio_path
audio_backend = audio_result.backend
target = "full"
elif retry_action and retry_action.get("regenerate") in {"audio", "image"}:
target = retry_action["regenerate"]
retry_strategy = retry_action.get("strategy")
retry_metric = retry_action.get("failed_metric")
if target == "audio" and retry_action.get("audio_prompt"):
audio_prompt = retry_action["audio_prompt"]
if target == "image" and retry_action.get("image_prompt"):
img_pool = self.image_gen.retrieve_top_k(
retry_action["image_prompt"],
k=8,
)
else:
target = decide_regeneration(
last_scores["msci"],
last_scores["st_i"],
last_scores["st_a"],
self.msci_threshold,
)
if not handled_regen and target == "image":
idx = min(attempt - 1, max(len(img_pool) - 1, 0))
image_path = img_pool[idx][0] if img_pool else image_path
elif not handled_regen and target == "audio":
audio_path = str(paths.audio_dir / f"audio_attempt{attempt}.wav")
audio_prompt_variant = audio_prompt + f" Intensity level: {attempt}."
audio_result = self.audio_gen.generate(audio_prompt_variant, audio_path)
audio_backend = audio_result.backend
elif not handled_regen and target == "text":
narrative = self.text_gen.generate(
f"{plan_text}\n\nRewrite concisely, keep the same meaning:\n"
).text
else:
target = "none"
if not image_path:
raise RuntimeError("Image path is empty; retrieval produced no candidates.")
image_emb = self.embedder.embed_image(image_path)
audio_emb = self.embedder.embed_audio(audio_path)
msci = compute_msci_v0(
plan_embedding,
image_emb,
audio_emb,
include_image_audio=True,
)
drift = detect_drift(msci.msci, msci.st_i, msci.st_a, msci.si_a)
scores = {
"msci": msci.msci,
"st_i": msci.st_i,
"st_a": msci.st_a,
"si_a": msci.si_a,
"agreement_score": merge_report["agreement_score"],
"per_section_agreement": merge_report["per_section_agreement"],
}
metric_scores = {k: scores[k] for k in ("msci", "st_i", "st_a", "si_a")}
coherence_step = self.coherence_scorer.score(
scores=metric_scores,
global_drift=drift["global_drift"],
)
coherence_step["needs_repair"] = (
coherence_step["classification"]["label"] == "MODALITY_FAILURE"
and coherence_step["classification"]["weakest_metric"] == "st_i"
)
repair_attempts = 0
while coherence_step["needs_repair"] and repair_attempts < 2:
narrative_structured = self.narrative_generator.repair_visual_description(
plan.model_dump(),
image_path=image_path,
)
narrative = narrative_structured.combined_scene
plan_embedding = self.embedder.embed_text(
narrative_structured.visual_description
)
msci = compute_msci_v0(
plan_embedding,
image_emb,
audio_emb,
include_image_audio=True,
)
drift = detect_drift(msci.msci, msci.st_i, msci.st_a, msci.si_a)
scores = {
"msci": msci.msci,
"st_i": msci.st_i,
"st_a": msci.st_a,
"si_a": msci.si_a,
"agreement_score": merge_report["agreement_score"],
"per_section_agreement": merge_report["per_section_agreement"],
}
metric_scores = {k: scores[k] for k in ("msci", "st_i", "st_a", "si_a")}
coherence_step = self.coherence_scorer.score(
scores=metric_scores,
global_drift=drift["global_drift"],
)
coherence_step["needs_repair"] = (
coherence_step["classification"]["label"] == "MODALITY_FAILURE"
and coherence_step["classification"]["weakest_metric"] == "st_i"
)
repair_attempts += 1
if coherence_step["classification"]["label"] in {
"HIGH_COHERENCE",
"LOCAL_MODALITY_WEAKNESS",
}:
break
step_decision = {
"attempt": attempt,
"image_path": image_path,
"audio_path": audio_path,
"audio_backend": audio_backend if "audio_backend" in locals() else None,
"scores": scores,
"coherence": coherence_step,
"drift": drift,
"retry_strategy": retry_strategy if attempt > 1 else None,
"retry_metric": retry_metric if attempt > 1 else None,
}
decisions.append(step_decision)
if attempt > 1 and retry_metric:
prev_scores = decisions[-2].get("scores", {})
before = prev_scores.get(retry_metric)
after = scores.get(retry_metric)
if before is not None and after is not None:
before_status = self.coherence_scorer.thresholds.classify_value(
retry_metric,
before,
)
after_status = self.coherence_scorer.thresholds.classify_value(
retry_metric,
after,
)
success = (before_status == "FAIL" and after_status in {"WEAK", "GOOD"}) or (
after > before + epsilon
)
retry_outcomes.append(
{
"strategy": retry_strategy,
"trigger": retry_trigger,
"weakest_metric": retry_metric,
"before": {
"msci": prev_scores.get("msci"),
"st_i": prev_scores.get("st_i"),
"st_a": prev_scores.get("st_a"),
"si_a": prev_scores.get("si_a"),
},
"after": {
"msci": scores.get("msci"),
"st_i": scores.get("st_i"),
"st_a": scores.get("st_a"),
"si_a": scores.get("si_a"),
},
"epsilon": epsilon,
"success": success,
}
)
if best_state is None or scores["msci"] > best_state[0]:
best_state = (
scores["msci"],
narrative,
image_path,
audio_path,
scores,
drift,
attempt,
)
if scores["msci"] >= self.msci_threshold and not drift["global_drift"]:
break
assert best_state is not None
_, best_text, best_img, best_aud, best_scores, best_drift, best_attempt = best_state
metric_scores = {k: best_scores[k] for k in ("msci", "st_i", "st_a", "si_a") if k in best_scores}
coherence = self.coherence_scorer.score(
scores=metric_scores,
global_drift=best_drift["global_drift"],
)
final_assessment = build_final_assessment(coherence, retry_outcomes)
out = RunOutput(
run_id=paths.run_id,
semantic_plan=plan.model_dump(),
merge_report=merge_report,
planner_outputs=planner_outputs,
narrative_structured=narrative_structured.model_dump(),
narrative_text=best_text,
image_path=best_img,
audio_path=best_aud,
scores=best_scores,
coherence=coherence,
final_assessment=final_assessment,
drift=best_drift,
attempts=best_attempt,
decisions=decisions,
)
write_run_metadata(
paths.logs_dir / "run.json",
{
"run_id": out.run_id,
"user_prompt": user_prompt,
"semantic_plan": out.semantic_plan,
"merge_report": out.merge_report,
"planner_outputs": out.planner_outputs,
"narrative_structured": out.narrative_structured,
"final": {
"narrative_text": out.narrative_text,
"image_path": out.image_path,
"audio_path": out.audio_path,
"scores": out.scores,
"coherence": out.coherence,
"final_assessment": out.final_assessment,
"drift": out.drift,
"attempts": out.attempts,
},
"attempt_history": out.decisions,
},
)
if retry_outcomes:
write_run_metadata(
paths.logs_dir / "retry_outcome.json",
{"retries": retry_outcomes},
)
return out
|