README.md

metadata

title: OpenEnv — Multi-Domain
emoji: 🌐
colorFrom: indigo
colorTo: purple
sdk: docker
pinned: false
license: apache-2.0
tags:
  - openenv
  - reinforcement-learning
  - agents
  - nlp
  - multi-domain
short_description: OpenEnv multi-domain evaluation environment for AI agents

OpenEnv — Multi-Domain AI Agent Environment

A comprehensive OpenEnv environment where AI agents are evaluated across 13 complex real-world tasks spanning 4 professional domains: Customer Support, Legal Review, Clinical Triage, and Software Engineering (PR Review).

---

🎯 Environment Description

To build general-purpose AI agents, we must evaluate them outside of traditional multiple-choice benchmarks. This environment provides 13 interactive, multi-step tasks across diverse verticals to test an agent's reasoning, tool use, and decision-making capabilities.

Agents must classify data, identify risks, draft domain-specific responses, and manage queues under dynamic constraints.

---

📋 Tasks & Domains

This environment features 13 tasks across 4 distinct professional domains:

🎫 Customer Support

Task	ID	Difficulty	Max Steps
Ticket Classification	ticket_classification	🟢 Easy	10
Response Drafting	response_drafting	🟡 Medium	6
SLA Queue Management	queue_management	🔴 Hard	40
Multi-Turn De-escalation	multi_turn_conversation	🟣 Very Hard	8

⚖️ Legal Review

Task	ID	Difficulty	Max Steps
Clause Identification	legal_clause_identification	🟢 Easy	10
Risk Flagging	legal_risk_flagging	🟡 Medium	10
Clause Redlining	legal_clause_redlining	🔴 Hard	5

🏥 Clinical Triage

Task	ID	Difficulty	Max Steps
Body System Classification	clinical_triage_classification	🟢 Easy	10
ESI Level Assignment	clinical_esi_assignment	🟡 Medium	10
Triage Note Generation	clinical_triage_note	🔴 Hard	5

💻 PR Review (Software Engineering)

Task	ID	Difficulty	Max Steps
PR Type Classification	pr_type_classification	🟢 Easy	10
Bug Identification	pr_bug_identification	🟡 Medium	5
Code Review Comment	pr_review_comment	🔴 Hard	5

---

🚀 Setup & Usage

Local Development

git clone https://huggingface.co/spaces/your-username/openenv-support
cd openenv-support

pip install -r requirements.txt

uvicorn app.main:app --host 0.0.0.0 --port 7860 --reload

Visit http://localhost:7860 for the interactive dashboard. Visit http://localhost:7860/docs for the Swagger API.

Docker

docker build -t openenv-support .
docker run -p 7860:7860 openenv-support

Running the Baseline

export OPENENV_BASE_URL="http://localhost:7860"

# Evaluate heuristic baseline on all tasks
python inference.py              

OpenEnv Validate

openenv validate --url http://localhost:7860

---

🌐 API Reference

Endpoint	Method	Description
/reset	POST	Reset episode (query param: task_id)
/step	POST	Take action, get observation + reward
/state	GET	Current state without advancing
/tasks	GET	List tasks with action schemas
/grader	POST	Get episode grader score
/baseline	POST	Run heuristic baseline on all tasks
/health	GET	Health check
/docs	GET	Swagger UI

---

📁 Project Structure

openenv-support/
├── app/
│   ├── main.py              # FastAPI application
│   └── models.py            # Typed Pydantic models
├── tasks/
│   ├── task1_classification.py
│   ├── task2_drafting.py
│   ├── task3_queue.py
│   ├── task4_multiturn.py
│   ├── legal_tasks.py
│   ├── clinical_tasks.py
│   └── pr_tasks.py
├── graders/
│   └── baseline_agent.py    # Auto-grader heuristics
├── data/
│   ├── tickets.py
│   ├── knowledge_base.py
│   ├── legal_data.py
│   ├── clinical_data.py
│   └── pr_data.py
├── static/
│   └── index.html           # Interactive multi-domain dashboard
├── openenv.yaml             # OpenEnv metadata
├── inference.py    # Automated test runner for baseline
├── validate.py              # Validation script
├── requirements.txt
├── Dockerfile
└── README.md

---

🏗️ Design Decisions

Why multi-domain? Real-world agents will need to operate across various specific verticals. Providing environments in Support, Legal, Clinical, and Engineering proves that an agent architecture is truly generalizable and robust.

Why deterministic graders? The rubric uses keyword matching, heuristic NLP signals (via `sentence-transformers`), semantic similarities, and exact comparisons — minimizing LLM-in-the-loop dependencies for grading and ensuring reproducible scores across runs for all 13 tasks.

Dynamic Reward Shaping: Every step returns a non-zero reward signal. Even incorrect actions return small penalties rather than zero, giving RL agents a gradient to follow rather than a sparse reward landscape.