openenv.yaml

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