iphone12man/email-prioritization-env

Smart Email Prioritization System

An OpenEnv-compatible reinforcement learning environment for real-world email triage. It models how an agent prioritizes emails, extracts deadlines, classifies categories, and drafts suggested replies.

Features

• —Priority classification (high, medium, low)
• —Deadline extraction (normalized datetime output)
• —Email categorization (work, college, personal, spam)
• —Reply suggestion generation

Task Levels

• —easy: priority only
• —medium: deadline + category
• —hard: priority + deadline + category + reply

Observation Space

Each observation contains:

• —episode_id
• —task_type
• —current_time
• —emails[], where each email includes:
• —id
• —sender
• —subject
• —body
• —timestamp

Action Space

The agent outputs:

• —email_id
• —priority
• —deadline (ISO-8601 or null)
• —category
• —suggested_reply

Actions are sent as:

{
  "actions": [
    {
      "email_id": "e1",
      "priority": "high",
      "deadline": null,
      "category": "work",
      "suggested_reply": "I will handle this and update you soon."
    }
  ]
}

Reward Function

Reward is task-aware and deterministic:

• —Component scoring:
• —priority exact match
• —deadline exact/same-date/incorrect scoring
• —category exact match
• —reply keyword and length checks
• —Penalties:
• —missing actions
• —extra actions
• —invalid values
• —missing or hallucinated deadlines
• —Final reward is normalized to [0.0, 1.0].

Project Structure

email_env/
├── env/
│   ├── environment.py
│   ├── grader.py
├── inference.py
├── requirements.txt
├── Dockerfile
├── openenv.yaml
├── README.md

Setup (Docker)

docker build -t email-env .
docker run email-env

Inference Usage

python inference.py

Notes

• —Uses OpenAI client with deterministic fallback logic when API is unavailable.
• —Evaluation pipeline is deterministic for stable benchmark behavior.