Apppublic

f4waz/InboxPilot

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

InboxPilot

InboxPilot is a real-world OpenEnv-compatible AI environment for email triage. The environment simulates a human inbox workflow where an AI agent must classify incoming emails, prioritize them, draft safe and useful replies, and avoid incorrect actions.

Overview

—Lightweight: Uses a clean Python-only implementation with a FastAPI wrapper.
—Tasks:
—Easy: spam / non-spam classification
—Medium: classify + draft reply
—Hard: full multi-step inbox handling with prioritization, categorization, and reply generation
—Scoring: Deterministic grader returning a StepReward containing reward, progress_score, and explanation.

Project Structure

—app.py: FastAPI server wrapper exposing OpenEnv-style generic endpoints.
—env.py: Contains InboxPilotEnv logic.
—models.py: Strongly-typed Pydantic schemas defining Observation, Action, StepReward, EnvironmentState.
—tasks.py: Definitions for Easy, Medium, and Hard task episodes.
—inference.py: Baseline agent execution loop using OpenAI's client.

API Endpoints

—GET / - Health check.
—POST /reset - Resets the env for a given task ID. Returns Observation.
—POST /step - Passes an Action to the env. Returns Observation, StepReward, is_done.
—GET /state - Access internal EnvironmentState.

Running the API

You can run the environment natively or via Docker.

Native

bash

pip install -r requirements.txt
uvicorn app:app --reload

Docker

bash

docker build -t inboxpilot .
docker run -p 8000:8000 inboxpilot

Running the Agent

Provide your OpenAI API key and start the API, then run the inference loop.

bash

export OPENAI_API_KEY="your-api-key"
# export API_BASE_URL="http://localhost:8000" (default)
python inference.py