File size: 5,272 Bytes
2930dae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a3c6db2
 
2930dae
a3c6db2
2930dae
 
 
 
 
 
 
 
 
 
a3c6db2
 
2930dae
a3c6db2
2930dae
 
 
 
 
 
 
 
 
 
 
a3c6db2
 
2930dae
 
 
a3c6db2
2930dae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
name: SupportEnv
version: "1.0.0"
description: >
  An OpenEnv-compliant customer support ticket triage environment for SaaS platforms.
  Agents learn to classify tickets, extract structured information, and generate
  professional resolutions — skills directly transferable to real-world support automation.

domain: customer_support
tags:
  - openenv
  - customer-support
  - nlp
  - information-extraction
  - classification
  - generation
  - real-world

license: MIT
author: SupportEnv Contributors

# -------------------------------------------------
# Environment interface
# -------------------------------------------------
interface:
  reset:
    method: POST
    path: /reset
    request:
      task_id: string        # task1 | task2 | task3
      ticket_index: integer  # optional, 0-4
    response: Observation

  step:
    method: POST
    path: /step
    request:
      episode_id: string
      action: Action
    response: StepResult     # {observation, reward, done, info}

  state:
    method: GET
    path: /state
    params:
      episode_id: string
    response: State

  tasks:
    method: GET
    path: /tasks
    response: List[TaskInfo]

  grader:
    method: POST
    path: /grader
    request:
      episode_id: string
    response: GraderResponse

  health:
    method: GET
    path: /health

# -------------------------------------------------
# Typed models
# -------------------------------------------------
models:
  Observation:
    task_id: string
    task_description: string
    episode_id: string
    ticket: TicketInfo
    thread_history: list[dict]
    available_actions: list[string]
    step_number: integer
    max_steps: integer
    hint: string | null

  TicketInfo:
    ticket_id: string
    subject: string
    body: string
    customer_tier: string    # free | pro | enterprise
    account_age_days: integer
    previous_tickets: integer
    attachments: list[string]

  Action:
    action_type: string      # classify | extract | respond | resolve | escalate | submit
    category: string | null
    priority: string | null
    extracted_entities: dict | null
    required_actions: list[string] | null
    response_text: string | null
    resolution_steps: list[string] | null
    escalation_team: string | null
    escalation_reason: string | null

  Reward:
    step_reward: float
    total_reward: float
    explanation: string

  StepResult:
    observation: Observation
    reward: Reward
    done: boolean
    info: dict

  State:
    task_id: string
    episode_id: string
    step_number: integer
    max_steps: integer
    done: boolean
    total_reward: float
    history: list[dict]
    final_score: float | null

  GraderResponse:
    episode_id: string
    task_id: string
    score: float              # 0.0 – 1.0
    breakdown: dict[string, float]
    feedback: string

# -------------------------------------------------
# Tasks
# -------------------------------------------------
tasks:
  task1:
    name: "Ticket Classification"
    difficulty: easy
    max_steps: 3
    description: >
      Given a customer support ticket, classify it by category
      (billing | technical | account | feature_request | complaint | general)
      and priority (low | medium | high | critical).
    scoring:
      baseline: 0.01
      category_correct: 0.49
      priority_correct: 0.40
      efficiency: 0.09
    tickets: 5

  task2:
    name: "Information Extraction"
    difficulty: medium
    max_steps: 5
    description: >
      Extract structured entities (account IDs, names, amounts, dates, domains)
      from the ticket body and identify the list of required actions.
    scoring:
      baseline: 0.01
      entity_coverage: 0.59
      action_coverage: 0.30
      no_hallucination: 0.09
    tickets: 5

  task3:
    name: "Resolution Generation"
    difficulty: hard
    max_steps: 8
    description: >
      Generate a professional customer-facing response (response_text) and
      an ordered list of resolution steps. Scored on keyword coverage,
      step completeness, tone (apology, urgency, timeline), and response length.
    scoring:
      baseline: 0.01
      keyword_coverage: 0.29
      step_coverage: 0.30
      tone_compliance: 0.25
      length_adequate: 0.10
      no_empty_steps: 0.04
    tickets: 5

# -------------------------------------------------
# Reward design
# -------------------------------------------------
reward:
  type: dense
  step_cost: -0.02       # small cost per step (encourages efficiency)
  submit_bonus: 0.05     # bonus for explicit submit action
  max_step_penalty: -0.10 # penalty for exhausting max_steps
  grader_bonus: up_to_1.0 # grader score (0–1) added as terminal bonus

# -------------------------------------------------
# Reproducibility
# -------------------------------------------------
reproducibility:
  dataset: static  # all 15 tickets are fixed, no randomisation
  graders: deterministic  # rule-based, no LLM judge
  baseline_mode: heuristic  # no API key required for reference scores

# -------------------------------------------------
# Deployment
# -------------------------------------------------
deployment:
  framework: FastAPI
  python: ">=3.10"
  port: 7860
  dockerfile: Dockerfile
  huggingface_space: true
  space_sdk: docker