--- title: Document Classification Env emoji: 📄 colorFrom: blue colorTo: green sdk: docker app_file: app.py pinned: false --- # 🌌 Void Walkers: Intelligent Routing Environment **A premium, Gymnasium-compatible Document Classification Environment designed for AI Agentic Learning.** --- ## 🚀 Overview Void Walkers represents a state-of-the-art **Partially Observable Markov Decision Process (POMDP)** where agents learn to classify, route, and interact with complex customer support payloads. Unlike static datasets, this environment enforces real-world constraints: **SLA Time Pressure**, **Processing Costs**, and **Tool-Mediated Metadata Retrieval**. ### 🌟 Key Enhancements (v1.1) - **Synchronized Category Space**: Standardized 22 fine-grained categories across the environment, agent, and spec. - **Dynamic Reward Architecture**: Implemented real-time Speed Bonuses and Resource Costs in the `step()` loop. - **Rich State API**: Deep telemetry including `action_history` and `reward_history` for diagnostic observability. - **Live RL Training**: Integrated PyTorch-based Reinforcement Learning loop demonstrating agent improvement over time. --- ## 🛠️ System Architecture ### 1. Environment: `DocumentClassificationEnv` The core logic resides in `environment.py`, implementing the standard `step()`, `reset()`, and `state()` API. - **Obs Space**: Dictionary containing raw `content`, `features` (100d), `urgency`, and `sla_remaining`. - **Action Space**: Discrete(24) — 22 labels + 2 tools (`Request Metadata`, `Escalate`). - **Reward Function**: $$R = \text{Accuracy} + \text{Speed Bonus} - \text{SLA Penalty} - \text{Processing Cost}$$ ### 2. Agent: `TicketAgent` A hybrid intelligence system: - **Baseline**: TF-IDF + Logistic Regression (pre-trained). - **RL Agent**: A neural policy that learns to map features to categories based on environment feedback. --- ## 📈 Benchmarks & Learning | Metric | Target (Easy) | Target (Hard) | | :--- | :--- | :--- | | **Accuracy** | 95%+ | 30%+ (High Entropy) | | **SLA Compliance** | 100% | >85% | | **Latency** | <50ms | <1.0s | --- ## 📖 Quick Start ### 1. Initialize Sandbox ```python from environment import DocumentClassificationEnv env = DocumentClassificationEnv(task_difficulty="medium") obs, info = env.reset() ``` ### 2. Execute Action ```python action = 1 # "Billing" obs, reward, done, truncated, info = env.step(action) ``` ### 3. Observe State ```python telemetry = env.state() print(f"Total Reward: {telemetry['total_reward']}") ``` --- ## 🖥️ UI & Visualization Run the Gradio dashboard for interactive testing: ```bash python app.py ``` *Navigate to the **Reinforcement Learning** tab to witness the neural training sequence.* --- ## 📜 Documentation & Spec Full specifications are defined in [openenv.yaml](./openenv.yaml). This project adheres to the **OpenEnv v1.0 Standard**. **Developed for Advanced AI Agentic Coding Research.**