File size: 10,836 Bytes
bd67155 | 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 | from __future__ import annotations
from typing import Dict, List, Optional, Tuple
from .graders import grade_task
from .models import Action, Observation, RewardModel, StateModel, StepInfo, TaskSpec, TicketObservation
from .reward import STEP_PENALTY, build_reward
from .state import discovered_for_ticket, initial_tracking, update_mapping
from .tasks import get_all_tasks, get_task
class SupportOpsEnv:
"""OpenEnv-shaped benchmark for support operations workflows."""
def __init__(self, task_id: Optional[str] = None):
self._tasks = {task.task_id: task for task in get_all_tasks()}
self._task_order = sorted(self._tasks)
self._task_id = task_id or self._task_order[0]
self._task: TaskSpec = self._tasks[self._task_id]
self._state: StateModel = initial_tracking(self._task)
def reset(self, task_id: Optional[str] = None) -> Observation:
if task_id is not None:
self._task = get_task(task_id)
self._task_id = task_id
self._state = initial_tracking(self._task)
return self._build_observation()
def state(self) -> StateModel:
return self._state.model_copy(deep=True)
def step(self, action: Action) -> Tuple[Observation, RewardModel, bool, Dict[str, object]]:
if self._state.done:
reward = build_reward({"invalid_after_done": -0.1}, "Episode already finished.")
info = StepInfo(
task_id=self._task.task_id,
step_count=self._state.step_count,
task_score=self._state.latest_score.get("task_score", 0.0),
done_reason="already_done",
event="invalid_after_done",
event_score=reward.components,
)
return self._build_observation(), reward, True, info.model_dump()
self._state.step_count += 1
event_scores: Dict[str, float] = {"step_penalty": STEP_PENALTY}
event_name = action.action_type
done_reason = None
if action.action_type == "inspect_ticket":
event_scores.update(self._handle_inspect(action))
elif action.action_type == "request_context":
event_scores.update(self._handle_request_context(action))
elif action.action_type == "set_priority":
event_scores.update(self._handle_priority(action))
elif action.action_type == "set_route":
event_scores.update(self._handle_route(action))
elif action.action_type == "set_resolution":
event_scores.update(self._handle_resolution(action))
elif action.action_type == "escalate":
event_scores.update(self._handle_escalation(action))
elif action.action_type == "rank_queue":
event_scores.update(self._handle_rank_queue(action))
elif action.action_type == "finalize":
self._state.done = True
done_reason = "agent_finalized"
grade = grade_task(self._task, self._state)
self._state.latest_score = {"task_score": grade.score, **grade.component_scores}
event_scores["terminal_grade"] = grade.score
reward = build_reward(event_scores, "Final task grade applied.")
self._state.cumulative_reward = round(self._state.cumulative_reward + reward.value, 4)
info = StepInfo(
task_id=self._task.task_id,
step_count=self._state.step_count,
task_score=grade.score,
done_reason=done_reason,
grade=grade,
event=event_name,
event_score=reward.components,
)
return self._build_observation(), reward, True, info.model_dump()
else:
event_scores["invalid_action"] = -0.1
event_name = "invalid_action"
grade = grade_task(self._task, self._state)
self._state.latest_score = {"task_score": grade.score, **grade.component_scores}
if self._state.step_count >= self._task.max_steps and not self._state.done:
self._state.done = True
done_reason = "max_steps"
event_scores["timeout_grade"] = grade.score
reward = build_reward(event_scores, f"Processed {event_name}.")
self._state.cumulative_reward = round(self._state.cumulative_reward + reward.value, 4)
info = StepInfo(
task_id=self._task.task_id,
step_count=self._state.step_count,
task_score=grade.score,
done_reason=done_reason,
grade=grade if self._state.done else None,
event=event_name,
event_score=reward.components,
)
return self._build_observation(), reward, self._state.done, info.model_dump()
def _build_observation(self) -> Observation:
tickets: List[TicketObservation] = []
for ticket in self._task.tickets:
keys = self._state.discovered_keys.get(ticket.ticket_id, [])
discovered_context = {key: ticket.hidden_context[key] for key in keys}
tickets.append(
TicketObservation(
ticket_id=ticket.ticket_id,
summary=ticket.summary,
visible_context=ticket.visible_context,
discovered_context=discovered_context,
selected_priority=self._state.priorities.get(ticket.ticket_id),
selected_route=self._state.routes.get(ticket.ticket_id),
selected_resolution=self._state.resolutions.get(ticket.ticket_id),
escalation_team=self._state.escalations.get(ticket.ticket_id),
)
)
return Observation(
task_id=self._task.task_id,
difficulty=self._task.difficulty,
title=self._task.title,
instruction=self._task.instruction,
queue_mode=self._task.queue_mode,
tickets=tickets,
remaining_steps=max(self._task.max_steps - self._state.step_count, 0),
available_actions=[
"inspect_ticket",
"request_context",
"set_priority",
"set_route",
"set_resolution",
"escalate",
"rank_queue",
"finalize",
],
current_queue_order=self._state.queue_order,
score_hint=self._state.latest_score,
)
def _find_ticket(self, ticket_id: str):
for ticket in self._task.tickets:
if ticket.ticket_id == ticket_id:
return ticket
return None
def _handle_inspect(self, action: Action) -> Dict[str, float]:
ticket = self._find_ticket(action.target)
if ticket is None:
return {"invalid_ticket": -0.1}
key = f"inspected::{ticket.ticket_id}"
notes = self._state.latest_score.setdefault("inspections", 0.0)
if notes and key in self._state.latest_score:
return {"redundant_inspect": -0.03}
self._state.latest_score[key] = 1.0
return {"inspect": 0.03}
def _handle_request_context(self, action: Action) -> Dict[str, float]:
ticket = self._find_ticket(action.target)
if ticket is None or not action.value:
return {"invalid_context_request": -0.1}
if action.value not in ticket.hidden_context:
return {"unknown_context_key": -0.08}
discovered = discovered_for_ticket(self._state.discovered_keys, ticket.ticket_id)
if action.value in discovered:
return {"redundant_context_request": -0.05}
discovered.append(action.value)
if action.value in ticket.required_context:
return {"required_context_found": 0.12}
return {"optional_context_found": 0.04}
def _handle_priority(self, action: Action) -> Dict[str, float]:
ticket = self._find_ticket(action.target)
if ticket is None or not action.value:
return {"invalid_priority": -0.1}
current = self._state.priorities.get(ticket.ticket_id)
update_mapping(self._state.priorities, ticket.ticket_id, action.value)
if action.value == current:
return {"redundant_priority": -0.03}
return {"priority_match": 0.08 if action.value == ticket.gold_priority else -0.04}
def _handle_route(self, action: Action) -> Dict[str, float]:
ticket = self._find_ticket(action.target)
if ticket is None or not action.value:
return {"invalid_route": -0.1}
current = self._state.routes.get(ticket.ticket_id)
update_mapping(self._state.routes, ticket.ticket_id, action.value)
if action.value == current:
return {"redundant_route": -0.03}
return {"route_match": 0.1 if action.value == ticket.gold_route else -0.06}
def _handle_resolution(self, action: Action) -> Dict[str, float]:
ticket = self._find_ticket(action.target)
if ticket is None or not action.value:
return {"invalid_resolution": -0.1}
current = self._state.resolutions.get(ticket.ticket_id)
update_mapping(self._state.resolutions, ticket.ticket_id, action.value)
if action.value == current:
return {"redundant_resolution": -0.03}
return {"resolution_match": 0.12 if action.value == ticket.gold_resolution else -0.08}
def _handle_escalation(self, action: Action) -> Dict[str, float]:
ticket = self._find_ticket(action.target)
if ticket is None:
return {"invalid_escalation": -0.1}
team = action.value
current = self._state.escalations.get(ticket.ticket_id)
update_mapping(self._state.escalations, ticket.ticket_id, team)
if team == current:
return {"redundant_escalation": -0.03}
if team == ticket.gold_escalation_team:
return {"correct_escalation": 0.1}
if ticket.gold_escalation_team is None and team is None:
return {"correct_no_escalation": 0.03}
return {"incorrect_escalation": -0.1}
def _handle_rank_queue(self, action: Action) -> Dict[str, float]:
if not self._task.queue_mode or not action.value:
return {"invalid_queue_ranking": -0.1}
ranked = [item.strip() for item in action.value.split(",") if item.strip()]
valid_ticket_ids = {ticket.ticket_id for ticket in self._task.tickets}
if set(ranked) != valid_ticket_ids:
return {"malformed_queue_ranking": -0.08}
self._state.queue_order = ranked
correct_positions = sum(
1 for observed, expected in zip(ranked, self._task.gold_queue_order) if observed == expected
)
return {"queue_progress": round((correct_positions / len(ranked)) * 0.12, 4)}
|