README.md

title	Email Triage OpenEnv
emoji	📧
colorFrom	blue
colorTo	green
sdk	docker
pinned	false

# Email Triage Environment ## Problem Customer-support inboxes mix routine requests with urgent incidents. A triage system has to do more than classify text; it has to make usable decisions under constraints: - identify the issue type - assess urgency - choose an action (reply, escalate, archive) - keep behavior consistent across a full queue This project provides a repeatable environment to evaluate that behavior with task-level scoring. ## Approach The environment is split into three evaluation tasks that increase in difficulty. Each task runs through a full email queue and computes a score from step-level rewards. Design principles used here: - one shared environment contract for all tasks - stricter requirements as difficulty increases - deterministic baseline available at all times - optional LLM-backed path for harder behavior checks The result is practical: you can compare stable baseline behavior vs model-assisted behavior without changing the core environment loop. ## Tasks - task_easy: basic triage signal quality (mainly category correctness) - task_medium: category plus priority/action quality - task_hard: category , priority , action , reply_template ## Results (Current Snapshot) Recent runs are typically in this range: - task_easy: about 0.93 - task_medium: about 0.82 to 0.90 - task_hard: can vary more when model-assisted paths are active , currently for the dataset giving in range (0.76 - 0.79) Interpretation: - repeated identical scores usually mean deterministic path only - variability usually indicates model-assisted path or stochastic fallback behavior ## Setup 1. Create and activate your virtual environment. 2. Install dependencies: ```bash pip install -r requirements.txt ``` 3. Configure environment values in .env: ```dotenv API_BASE_URL=https://router.huggingface.co/v1 MODEL_NAME=Qwen/Qwen2.5-72B-Instruct HF_TOKEN=your_token_here ``` If provider calls fail (quota/rate limits), fallback logic is used where implemented. ## Running Run the evaluator: ```bash python inference.py ``` This executes easy, medium, and hard, printing START/STEP/END logs and one final summary line. Run the API server: ```bash python app.py ``` Default bind: - host: 0.0.0.0 - port: 7860 ## How To Check Use this API sequence: 1. POST /reset 2. POST /step (repeat until done=true) 3. GET /state 4. GET /health Expected behavior: - /reset initializes environment state and returns `{ "observation": { "current_email": { "id", "subject", "body" }, "remaining_emails": N } }` - /step processes exactly one email using `{ "action": "classify_email" | "extract_entities" | "generate_reply" }` and returns per-step reward - /state returns current index and done flag for the active episode ## Notes - Keep .env and tokens private. - For stable benchmarking, use deterministic mode. - For behavior stress tests, enable model-assisted paths. # force redeploy 04/12/2026 18:10:14 Final deployment trigger - Random_Rugs