AutoDataLab2.0 / ceo_brief_env /environment.py
uchihamadara1816's picture
Upload 172 files
d02bacd verified
from __future__ import annotations
import json
from pathlib import Path
from typing import Any, Dict
from uuid import uuid4
import pandas as pd
from .experts import DataAnalystExpert, FinanceExpert, HRExpert, StrategyExpert
from .graders import grade_episode, load_metric_ground_truth
from .models import Brief, CoSAction, CoSObservation, CoSState, ExpertReport, RewardBreakdown
TASK_ROOT = Path(__file__).resolve().parent / 'tasks'
# Per-task order for “single” baselines, oracle, and any policy that should cover required experts.
REQUIRED_EXPERTS_BY_TASK: dict[str, list[str]] = {
'easy_brief': ['analyst', 'finance', 'hr'],
'medium_brief': ['analyst', 'finance', 'strategy', 'hr'],
'hard_brief': ['analyst', 'finance', 'strategy', 'hr'],
'expert_brief': ['analyst', 'finance', 'strategy', 'hr'],
'risk_brief': ['analyst', 'finance', 'strategy', 'hr'],
'crisis_brief': ['analyst', 'finance', 'strategy', 'hr'],
}
def required_experts_for_task(task_name: str) -> list[str]:
return list(REQUIRED_EXPERTS_BY_TASK.get(task_name, ['analyst', 'finance', 'hr']))
class CEOBriefEnvironment:
def __init__(self, shaping: str = "default", auto_fill_required: bool = True) -> None:
"""Multi-agent CEO-brief env.
``shaping`` controls the dense per-step reward. The terminal grader is
unchanged either way; that is what hackathon scoring uses.
- ``"default"``: legacy per-step rewards. Stable; matches existing
trained checkpoints and submitted runs.
- ``"strict"``: anti-degenerate shaping for RL training. Adds a
repetition penalty, an over-consult penalty, an early-finish bonus
when all required experts are covered, and a stronger penalty for
summarizing before required experts have reported. Use for new
GRPO/REINFORCE runs to discourage "summarize-spam -> submit" lazy
policies.
``auto_fill_required`` keeps the production/demo environment robust by
filling any missing required experts before composing or grading. Turn
it off only for policy-evidence runs where we want to observe what the
LLM actually routed by itself.
"""
self.analyst = DataAnalystExpert()
self.finance = FinanceExpert()
self.hr = HRExpert()
self.strategy = StrategyExpert()
self.use_rag = False
self.shaping = shaping if shaping in {"default", "strict"} else "default"
self.auto_fill_required = auto_fill_required
self.reset()
def reset(self, task: str = 'easy_brief', episode_id: str | None = None, use_rag: bool = False) -> CoSObservation:
self.use_rag = use_rag
self.episode_id = episode_id or str(uuid4())
self.task_name = task if (TASK_ROOT / task).exists() else 'easy_brief'
task_dir = TASK_ROOT / self.task_name
self.raw_df = pd.read_csv(task_dir / 'raw.csv')
self.gt_metrics = load_metric_ground_truth(str(task_dir / 'ground_truth.csv')) if (task_dir / 'ground_truth.csv').exists() else {}
with open(task_dir / 'metadata.json', encoding='utf-8') as f:
self.meta = json.load(f)
self.step_count = 0
self.done = False
self.cumulative_reward = 0.0
self.expert_reports: Dict[str, ExpertReport] = {}
self.current_brief: Brief | None = None
self.history: list[str] = []
self.last_reward = 0.0
self.last_terminal = None
self.last_data_quality = 0.0
self.last_issues = ['No experts consulted yet.']
self._consult_counts: Dict[str, int] = {}
self._last_action_key: str | None = None
return self._observe(initial=True)
def state(self) -> CoSState:
return CoSState(
episode_id=self.episode_id,
task_name=self.task_name,
step_count=self.step_count,
done=self.done,
rag_enabled=self.use_rag,
consulted_experts=list(self.expert_reports.keys()),
expert_reports=self.expert_reports,
current_brief=self.current_brief,
cumulative_reward=self.cumulative_reward,
)
def _observe(self, initial: bool = False) -> CoSObservation:
return CoSObservation(
done=self.done,
reward=0.0 if initial else self.last_reward,
instruction=self.meta['instruction'],
history=list(self.history),
issues=list(self.last_issues),
data_quality_score=self.last_data_quality,
task_name=self.task_name,
task_difficulty=self.meta['difficulty'],
max_steps=int(self.meta.get('max_steps', 12)),
step_count=self.step_count,
rag_enabled=self.use_rag,
consulted_experts=list(self.expert_reports.keys()),
expert_reports=self.expert_reports,
current_brief=self.current_brief,
reward_breakdown=RewardBreakdown(
immediate=self.last_reward,
cumulative=self.cumulative_reward,
terminal_grader=self.last_terminal,
),
terminal_grader_score=self.last_terminal,
)
def _compose_brief(self) -> Brief:
metrics: Dict[str, Any] = {}
recommendations: list[str] = []
summary_parts: list[str] = []
for expert_id in ('analyst', 'finance'):
report = self.expert_reports.get(expert_id)
if report:
metrics.update(report.metrics)
summary_parts.append(report.summary)
if 'strategy' in self.expert_reports:
recommendations = list(self.expert_reports['strategy'].bullet_points)
summary_parts.append(self.expert_reports['strategy'].summary)
hr_memo = self.expert_reports['hr'].memo if 'hr' in self.expert_reports and self.expert_reports['hr'].memo else ''
summary = ' '.join(summary_parts) if summary_parts else 'No brief drafted yet.'
self.current_brief = Brief(
summary=summary,
metrics=metrics,
recommendations=recommendations,
hr_memo=hr_memo,
consulted_experts=list(self.expert_reports.keys()),
)
return self.current_brief
def _run_expert(self, expert_id: str, focused: bool = False) -> ExpertReport:
question = self.meta['instruction']
if expert_id == 'analyst':
report = self.analyst.run(
self.task_name, question, self.raw_df, focused=focused, use_rag=self.use_rag
)
self.last_data_quality = float(report.metrics.get('data_quality_score', 0.0))
self.last_issues = report.issues or ['analyst:no material issues']
return report
if expert_id == 'finance':
analyst = self.expert_reports.get('analyst') or self._run_expert('analyst')
return self.finance.run(
self.task_name,
question,
self.raw_df,
analyst.metrics,
self.meta,
focused=focused,
use_rag=self.use_rag,
)
if expert_id == 'strategy':
analyst = self.expert_reports.get('analyst') or self._run_expert('analyst')
finance = self.expert_reports.get('finance') or self._run_expert('finance')
return self.strategy.run(
self.task_name, self.meta, analyst, finance, focused=focused, use_rag=self.use_rag
)
if expert_id == 'hr':
analyst = self.expert_reports.get('analyst') or self._run_expert('analyst')
finance = self.expert_reports.get('finance') or self._run_expert('finance')
strategy = self.expert_reports.get('strategy')
return self.hr.run(
self.task_name, self.meta, analyst, finance, strategy, focused=focused, use_rag=self.use_rag
)
raise ValueError(f'Unknown expert {expert_id!r}')
def _ensure_required_experts(self) -> list[str]:
"""Run any task-required experts that the policy never consulted.
This guarantees the strategist (and any other required role) always
contributes to the brief, so the UI / grader always has their report.
Returns the list of expert ids that were auto-filled.
"""
if not self.auto_fill_required:
return []
auto: list[str] = []
for expert_id in required_experts_for_task(self.task_name):
if expert_id in self.expert_reports:
continue
try:
self.expert_reports[expert_id] = self._run_expert(expert_id)
auto.append(expert_id)
except Exception:
continue
return auto
def step(self, action: CoSAction) -> CoSObservation:
if self.done:
return self._observe()
self.step_count += 1
immediate = -0.02
details = action.model_dump(exclude_none=True)
action_key = json.dumps(details, sort_keys=True)
self.history.append(action_key)
strict = self.shaping == 'strict'
if strict and self._last_action_key is not None and action_key == self._last_action_key:
immediate -= 0.05
self._last_action_key = action_key
required = list(self.meta.get('required_experts', []))
if action.action_type in {'consult', 'ask'}:
if not action.expert_id:
immediate -= 0.03
self.last_issues = ['action_missing_expert']
else:
prior = action.expert_id in self.expert_reports
report = self._run_expert(action.expert_id, focused=action.action_type == 'ask')
self.expert_reports[action.expert_id] = report
immediate += 0.10 if not prior and action.expert_id in required else 0.02
if prior:
immediate -= 0.05
if strict:
self._consult_counts[action.expert_id] = self._consult_counts.get(action.expert_id, 0) + 1
if self._consult_counts[action.expert_id] > 2:
immediate -= 0.10
self.last_issues = report.issues or [f'{action.expert_id}:ok']
elif action.action_type == 'summarize':
brief_already_exists = self.current_brief is not None
missing_required = [e for e in required if e not in self.expert_reports]
self._ensure_required_experts()
self._compose_brief()
immediate += 0.04 if len(self.expert_reports) >= 2 else -0.02
if strict and missing_required:
immediate -= 0.05 * len(missing_required)
if strict and brief_already_exists:
immediate -= 0.08
self.last_issues = ['brief_composed']
elif action.action_type == 'submit':
auto_filled = self._ensure_required_experts()
if self.current_brief is None or auto_filled:
self._compose_brief()
self.done = True
self.last_terminal = grade_episode(
self.gt_metrics, self.meta, self.current_brief, self.expert_reports, use_rag=self.use_rag
)
immediate += self.last_terminal
if strict and not auto_filled:
max_steps = int(self.meta.get('max_steps', 12))
steps_saved = max(0, max_steps - self.step_count)
if steps_saved > 0 and all(e in self.expert_reports for e in required):
immediate += min(0.10, 0.01 * steps_saved)
self.last_issues = ['submitted'] + (
[f'auto_consulted:{",".join(auto_filled)}'] if auto_filled else []
)
else:
self.last_issues = ['noop']
immediate -= 0.01
if not self.done and self.step_count >= int(self.meta.get('max_steps', 12)):
auto_filled = self._ensure_required_experts()
if self.current_brief is None or auto_filled:
self._compose_brief()
self.done = True
self.last_terminal = grade_episode(
self.gt_metrics, self.meta, self.current_brief, self.expert_reports, use_rag=self.use_rag
)
immediate += self.last_terminal
self.last_issues = ['forced_termination:max_steps'] + (
[f'auto_consulted:{",".join(auto_filled)}'] if auto_filled else []
)
self.last_reward = round(immediate, 4)
self.cumulative_reward = round(self.cumulative_reward + self.last_reward, 4)
return self._observe()
def oracle_action_for_observation(obs: CoSObservation) -> CoSAction:
for expert in required_experts_for_task(obs.task_name):
if expert not in obs.consulted_experts:
return CoSAction(action_type='consult', expert_id=expert)
if obs.current_brief is None:
return CoSAction(action_type='summarize')
return CoSAction(action_type='submit')