Spaces:
Sleeping
Sleeping
File size: 9,624 Bytes
ce8f04a | 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 | """
Expectation quality loop: evaluate → targets → re-evaluate until ready or max-iter.
Pure control plane (no LLM). Callers inject evaluate/fix callables so unit tests
drive the real loop without network.
"""
from __future__ import annotations
import logging
import os
from dataclasses import asdict, dataclass, field
from typing import Any, Callable, Dict, List, Optional
logger = logging.getLogger(__name__)
DEFAULT_MAX_ITER = int(os.environ.get("ADVISOR_EXPECTATION_MAX_ITER", "3"))
DEFAULT_MIN_SCORE = int(os.environ.get("ADVISOR_EXPECTATION_MIN_SCORE", "70"))
@dataclass
class ExpectationLoopResult:
ready: bool
iterations: int
max_iterations: int
final_score: int
regulation_grounded_pass: bool
stop_reason: str # ready | max_iter | blocked_grounding | evaluate_error
history: List[Dict[str, Any]] = field(default_factory=list)
remaining_blockers: List[str] = field(default_factory=list)
final_report: Optional[Dict[str, Any]] = None
state: Optional[Dict[str, Any]] = None
def to_dict(self) -> Dict[str, Any]:
return asdict(self)
def is_regulation_ready(
report: Any,
*,
min_score: int = DEFAULT_MIN_SCORE,
require_regulation_grounded: bool = True,
) -> bool:
"""Stop condition: advisor report meets readiness criteria."""
if report is None:
return False
if isinstance(report, dict):
score = int(report.get("score") or 0)
grounded = bool(report.get("regulation_grounded_pass"))
passed = bool(report.get("passed"))
blockers = list(report.get("blockers") or [])
mode = str(report.get("grounding_mode") or "regulation")
else:
score = int(getattr(report, "score", 0) or 0)
grounded = bool(getattr(report, "regulation_grounded_pass", False))
passed = bool(getattr(report, "passed", False))
blockers = list(getattr(report, "blockers", None) or [])
mode = str(getattr(report, "grounding_mode", "regulation") or "regulation")
if mode in ("structure_only", "blocked", "blind"):
# Never regulation-ready in ungrounded modes
return False
if require_regulation_grounded and not grounded:
return False
if blockers:
return False
if score < min_score:
return False
return passed or grounded
def extract_blockers(report: Any) -> List[str]:
if report is None:
return ["brak raportu doradcy"]
if isinstance(report, dict):
blockers = list(report.get("blockers") or [])
if blockers:
return [str(b) for b in blockers]
findings = report.get("findings") or []
out = []
for f in findings:
if isinstance(f, dict) and (f.get("blocking") or f.get("severity") == "critical"):
out.append(str(f.get("message") or f.get("code") or "finding"))
return out
blockers = list(getattr(report, "blockers", None) or [])
if blockers:
return [str(b) for b in blockers]
out = []
for f in getattr(report, "findings", None) or []:
if getattr(f, "blocking", False) or getattr(f, "severity", "") == "critical":
out.append(str(getattr(f, "message", "") or getattr(f, "code", "finding")))
return out
def report_to_dict(report: Any) -> Dict[str, Any]:
if report is None:
return {}
if isinstance(report, dict):
return report
if hasattr(report, "to_dict"):
return report.to_dict()
return {
"passed": getattr(report, "passed", False),
"score": getattr(report, "score", 0),
"regulation_grounded_pass": getattr(report, "regulation_grounded_pass", False),
"blockers": list(getattr(report, "blockers", None) or []),
"grounding_mode": getattr(report, "grounding_mode", ""),
"summary": getattr(report, "summary", ""),
}
def run_expectation_loop(
*,
evaluate: Callable[[Dict[str, Any]], Any],
apply_fixes: Optional[Callable[[Dict[str, Any], Any], Dict[str, Any]]] = None,
initial_state: Optional[Dict[str, Any]] = None,
max_iterations: int = DEFAULT_MAX_ITER,
min_score: int = DEFAULT_MIN_SCORE,
require_regulation_grounded: bool = True,
) -> ExpectationLoopResult:
"""
Loop:
evaluate(state) → if ready: stop
else apply_fixes(state, report) → state' → re-evaluate
until ready or max_iterations.
structure_only / blind never stop as regulation-ready (is_regulation_ready=False).
"""
state: Dict[str, Any] = dict(initial_state or {})
max_iterations = max(1, int(max_iterations))
history: List[Dict[str, Any]] = []
last_report: Any = None
for i in range(1, max_iterations + 1):
try:
last_report = evaluate(state)
except Exception as e:
logger.warning("[ExpectationLoop] evaluate failed at iter %s: %s", i, e)
return ExpectationLoopResult(
ready=False,
iterations=i,
max_iterations=max_iterations,
final_score=0,
regulation_grounded_pass=False,
stop_reason="evaluate_error",
history=history,
remaining_blockers=[str(e)],
final_report={"error": str(e)},
state=state,
)
rd = report_to_dict(last_report)
history.append({"iteration": i, "phase": "evaluate", "report": rd})
mode = str(rd.get("grounding_mode") or state.get("grounding_mode") or "").lower()
ext = state.get("external_context") if isinstance(state.get("external_context"), dict) else {}
if not mode:
mode = str(ext.get("grounding_mode") or "regulation").lower()
rd["grounding_mode"] = mode
if mode in ("structure_only", "blocked", "blind"):
# Explicit: never mark regulation pass; stop early with blockers if blocked
if mode == "blocked":
return ExpectationLoopResult(
ready=False,
iterations=i,
max_iterations=max_iterations,
final_score=int(rd.get("score") or 0),
regulation_grounded_pass=False,
stop_reason="blocked_grounding",
history=history,
remaining_blockers=extract_blockers(last_report)
or ["Generacja zablokowana — brak regulaminu i zgody."],
final_report=rd,
state=state,
)
# structure_only: continue loop for structural quality but never ready=True for regulation
# Fall through to fix attempts, but is_regulation_ready stays False
if is_regulation_ready(
rd,
min_score=min_score,
require_regulation_grounded=require_regulation_grounded,
):
return ExpectationLoopResult(
ready=True,
iterations=i,
max_iterations=max_iterations,
final_score=int(rd.get("score") or 0),
regulation_grounded_pass=bool(rd.get("regulation_grounded_pass")),
stop_reason="ready",
history=history,
remaining_blockers=[],
final_report=rd,
state=state,
)
if i >= max_iterations:
break
if apply_fixes is None:
# No fixer → cannot improve; stop next iteration boundary
continue
try:
new_state = apply_fixes(state, last_report)
if isinstance(new_state, dict):
state = new_state
history.append(
{
"iteration": i,
"phase": "fix",
"fixed_sections": list(
(new_state or {}).get("fixed_sections")
or state.get("last_fixed_sections")
or []
)
if isinstance(new_state, dict)
else [],
}
)
except Exception as e:
logger.warning("[ExpectationLoop] apply_fixes failed at iter %s: %s", i, e)
history.append({"iteration": i, "phase": "fix_error", "error": str(e)})
rd = report_to_dict(last_report)
return ExpectationLoopResult(
ready=False,
iterations=max_iterations,
max_iterations=max_iterations,
final_score=int(rd.get("score") or 0),
regulation_grounded_pass=bool(rd.get("regulation_grounded_pass")),
stop_reason="max_iter",
history=history,
remaining_blockers=extract_blockers(last_report),
final_report=rd,
state=state,
)
def run_advisor_expectation_loop(
state: Dict[str, Any],
*,
max_iterations: int = DEFAULT_MAX_ITER,
min_score: int = DEFAULT_MIN_SCORE,
apply_fixes: Optional[Callable[[Dict[str, Any], Any], Dict[str, Any]]] = None,
) -> ExpectationLoopResult:
"""
Default loop using world_class_advisor.evaluate_from_generator_state.
Optional apply_fixes for rewrite integration (LLM or pure test stub).
"""
from agents.world_class_advisor import evaluate_from_generator_state
def _eval(st: Dict[str, Any]):
return evaluate_from_generator_state(st)
return run_expectation_loop(
evaluate=_eval,
apply_fixes=apply_fixes,
initial_state=state,
max_iterations=max_iterations,
min_score=min_score,
require_regulation_grounded=True,
)
|