openenv.yaml

name: CustomerSupportEnv
version: "1.0.0"
description: >
  A real-world customer support reinforcement learning environment where an AI agent
  handles inbound support tickets. The agent must search a knowledge base, empathise
  with customers, offer concrete solutions, and resolve tickets efficiently.
  Models the genuine complexity of Tier-1 customer support: multi-turn conversation,
  retrieval-augmented reasoning, and satisfaction optimisation.

author: OpenEnv Submission
domain: customer-support
tags: [openenv, customer-support, nlp, retrieval, multi-turn, real-world]

tasks:
  - id: task_1
    name: "Resolve a Standard Auth Ticket"
    difficulty: easy
    ticket: TKT-001
    max_turns: 8
    description: >
      Handle a frustrated customer locked out of their account.
      Optimal policy: search_kb → empathize → offer_solution → resolve.

  - id: task_2
    name: "Handle a Multi-Step Billing Dispute"
    difficulty: medium
    ticket: TKT-003
    max_turns: 10
    description: >
      Resolve a billing discrepancy. Requires clarification before diagnosis.
      Generic solutions are penalised; agent must cite a specific credit amount.

  - id: task_3
    name: "Triage a Critical Time-Sensitive Bug"
    difficulty: hard
    ticket: TKT-006
    max_turns: 8
    description: >
      Enterprise customer with a compliance deadline. Data export stuck for 6 hours.
      Two-part solution required (priority queue + partial export).
      Escalation is penalised. Tests urgency awareness and multi-step planning.

observation_space:
  type: object
  fields:
    ticket_id: {type: string, nullable: true}
    task_id: {type: string}
    status: {type: string, enum: [idle, open, resolved, escalated, timeout]}
    sentiment: {type: string, enum: [positive, neutral, frustrated, angry], nullable: true}
    priority: {type: string, enum: [low, medium, high, urgent], nullable: true}
    category: {type: string, enum: [auth, billing, fulfillment, bug, sales, general], nullable: true}
    turn: {type: integer, minimum: 0}
    max_turns: {type: integer}
    history: {type: array, items: {role: string, text: string, turn: integer}}
    kb_results: {type: array, items: {type: string}}
    kb_searched: {type: boolean}
    empathized: {type: boolean}
    clarified: {type: boolean}
    solution_offered: {type: boolean}
    escalated: {type: boolean}
    cumulative_reward: {type: number}
    done: {type: boolean}

action_space:
  type: object
  fields:
    action_type:
      type: string
      enum: [search_kb, empathize, ask_clarify, offer_solution, escalate, resolve, send_message]
    payload:
      type: string
      nullable: true
      description: >
        Required for offer_solution (solution text), ask_clarify (question),
        and send_message (message body). Optional for others.

reward_function:
  type: shaped
  components:
    search_kb: "+2.0 (first call only; -1.0 duplicate)"
    empathize: "+1.0 (first call only)"
    ask_clarify: "+1.0 (first call only)"
    offer_solution: "+3.0 × quality_score (0–1); -1.0 if KB not searched first"
    escalate: "-1.0"
    resolve_good: "+5.0 + csat × 2.0 (when solution offered)"
    resolve_bad: "-3.0 (when no solution offered)"
    timeout: "-2.0"
  csat_components:
    empathized: 0.30
    kb_searched: 0.30
    solution_offered: 0.40

graders:
  scoring: 0.0_to_1.0
  deterministic: true
  task_1_weights:
    kb_searched: 0.30
    empathized: 0.25
    solution_quality: 0.25
    resolved: 0.20
  task_2_weights:
    ask_clarify: 0.20
    kb_searched: 0.20
    solution_quality: 0.30
    empathized: 0.15
    resolved: 0.15
  task_3_weights:
    kb_searched: 0.20
    empathized: 0.15
    solution_quality: 0.35
    no_escalation: 0.15
    resolved: 0.15

endpoints:
  reset: "POST /reset"
  step: "POST /step"
  state: "GET /state"
  tasks: "GET /tasks"
  grade: "POST /grade"
  health: "GET /health"
  spec: "GET /openenv.yaml"

baseline_scores:
  task_1: 0.85
  task_2: 0.78
  task_3: 0.65
  average: 0.76
  model: gpt-4o-mini

huggingface:
  space_sdk: docker
  port: 7860