Docs: Simplify README to core technical specifications

README.md

@@ -1,85 +1,70 @@
-# 🚀 RAG Context Optimizer: Enterprise Incident Operations
+# RAG Context Optimizer
 
-**Context Optimizer** is an advanced Reinforcement Learning (RL) environment designed for enterprise-grade context management in RAG pipelines. It simulates the high-stakes decisions made during live operational incidents like outages, security escalations, and cross-functional briefings.
+A reinforcement learning environment for context optimization tasks including artifact inspection, prioritization, and summarization. This project simulates an operational workflow for handling structured data artifacts within token budgets.
 
----
+## Project Structure
 
-## 💡 Motivation: Operational Intelligence
-Standard RAG often fails in complex scenarios where picking the right context isn't just about semantic similarity. Incident commanders and support leads must:
-- **Inspect** artifacts across multiple domains (Support, Engineering, Billing).
-- **Prioritize** evidence within strict token budgets.
-- **Summarize** heavy technical documents without losing grounding.
-- **Plan** resolutions before submitting final grounded memos.
+- `rag_optimizer_env/`: Core environment logic and task definitions.
+- `rag_gc_env/`: (Legacy) Context pruning environment for Round 1.
+- `app.py`: FastAPI implementation of the environment server.
+- `inference.py`: Baseline inference script using an OpenAI-compatible interface.
+- `optimizer_ui.py`: Streamlit-based dashboard for manual task execution.
+- `validate.py`: Local validation script for environment compliance.
+- `data/`: Corpus and task dataset files.
 
-This environment models these behaviors to benchmark and train agents that are accurate, efficient, and operationally safe.
+## Environment Interface
 
----
+### Action Space
 
-## 🎮 Action & Observation Spaces
+The agent interacts with the environment using the following actions:
 
-### Action Space (RagAction)
-| Action Type | Parameters | Effect |
-| --- | --- | --- |
-| `inspect_artifact` | `artifact_id` | Review an artifact without committing it to the working set |
-| `prioritize_artifact` | `artifact_id` | Add a reviewed artifact to the working set |
-| `summarize_artifact` | `artifact_id`, `ratio` | Compress an artifact to reduce token cost |
-| `set_resolution_plan` | `plan` | Draft the operational plan before submission |
-| `submit_report` | `answer` | Submit the final grounded memo and end the episode |
+- `inspect_artifact`: Review metadata for a chunk without committing to the context.
+- `prioritize_artifact`: Add an artifact to the current context (working set).
+- `summarize_artifact`: Compress a prioritized artifact according to a specified ratio.
+- `set_resolution_plan`: Define an operational plan.
+- `submit_report`: Submit the final grounded answer.
 
-### Observation Space (RagObservation)
-- **case_summary**: Real-world case context and background.
-- **objective**: The specific deliverable the agent must produce.
-- **workflow_stage**: Current phase (`triage`, `analysis`, `resolution`).
-- **available_artifacts**: Summaries of all artifacts available for review.
-- **token_budget**: Strict limit on the total tokens allowed in the working set.
-- **progress_signals**: Partial performance metrics throughout the trajectory.
+### API Endpoints
 
----
+- `POST /reset`: Initialize an episode for a given task.
+- `POST /step`: Execute an action and return the observation, reward, and terminal state.
+- `GET /state`: View the current episode state.
+- `POST /optimize-step`: Request a suggested action from the baseline policy.
+- `POST /optimize-prompt`: Optimize input prompts for grounding and length.
 
-## 🏆 Task Definitions
+## Tasks
 
-| Task Name | Difficulty | Max Steps | Token Budget | Objective |
-| :--- | :--- | :--- | :--- | :--- |
-| `refund_triage_easy` | **Easy** | 7 | 850 | Triage refund requests after an outage. |
-| `cross_function_brief_medium` | **Medium** | 8 | 620 | Sync support, IR, and release controls. |
-| `executive_escalation_hard` | **Hard** | 10 | 360 | Draft a terse, high-stakes compromise note. |
+| Name | Difficulty | Steps | Token Budget |
+| :--- | :--- | :--- | :--- |
+| `refund_triage_easy` | Easy | 7 | 850 |
+| `cross_function_brief_medium` | Medium | 8 | 620 |
+| `executive_escalation_hard` | Hard | 10 | 360 |
 
----
+## Setup and Execution
 
-## 🛠️ API Endpoints
+1. **Install dependencies**:
+   ```bash
+   pip install -r requirements.txt
+   ```
 
-The environment exposes a robust FastAPI interface:
-- `POST /reset`: Initialize a new episode with a specific task.
-- `POST /step`: Execute an action and receive reward/observation.
-- `GET /state`: Retrieve the full current state of the environment.
-- `POST /optimize-step`: **[AI Helper]** Get a suggested action from a baseline policy.
-- `POST /optimize-prompt`: **[AI Helper]** Rewrite prompts to fit budgets while preserving grounding.
+2. **Run server**:
+   ```bash
+   python app.py
+   ```
 
----
+3. **Inference**:
+   ```bash
+   python inference.py
+   ```
 
-## 🚀 Getting Started
+4. **UI Dashboard**:
+   ```bash
+   streamlit run optimizer_ui.py
+   ```
 
-### Installation
-```bash
-pip install -r requirements.txt
-```
+5. **Validation**:
+   ```bash
+   python validate.py
+   ```
 
-### Execution
-Start the environment server:
-```bash
-python app.py
-```
-
-Launch the Streamlit dashboard for manual optimization:
-```bash
-streamlit run optimizer_ui.py
-```
-
-### Validation
-Ensure full compliance with the benchmark specifications:
-```bash
-python validate.py
-```
-
----
-*Built for the Meta x Scaler Hackathon 2026*
+Built for the Meta x Scaler Hackathon 2026.