fix

README.md

@@ -15,13 +15,36 @@ tags:
 
 # SQL Agent OpenEnv
 
+> **Origin:** This project is a port of [gepa-tuned-sql-agent](https://github.com/Ar9av/gepa-tuned-sql-agent) — originally built as a TypeScript/Next.js application — rewritten in Python (FastAPI + React/Vite) to match the OpenEnv format and deploy on Hugging Face Spaces.
+
 An OpenEnv-compliant RL environment for SQL generation, featuring:
 
 - **LinUCB contextual bandit** — selects repair strategies based on error context
 - **GEPA (Generative Evolutionary Prompt Adaptation)** — evolves the system prompt from failure patterns
-- **3 difficulty tiers** — easy / medium / hard benchmark tasks on a SQLite e-commerce schema
+- **Multi-turn repair loop** — LLM receives full failure history so each retry learns from the previous error
+- **3 difficulty tiers** — easy / medium / hard benchmark tasks on a built-in e-commerce schema
 - **Shaped reward function** — success bonus, attempt penalty, error severity signal
-- **HER (Hindsight Experience Replay)** — relabels failed episodes for richer signal
+- **External database support** — connect any SQLite file or PostgreSQL database via connection string
+
+## Background
+
+The original [gepa-tuned-sql-agent](https://github.com/Ar9av/gepa-tuned-sql-agent) explored three research ideas in a Next.js stack:
+
+1. **Self-debug loop** — the agent critiques and fixes its own SQL errors without human intervention
+2. **GEPA prompt evolution** — after user feedback, an LLM reflects on failures and evolves the system prompt
+3. **Mini-RL environment** — a LinUCB contextual bandit learns which repair strategy works best for each error class
+
+This repo preserves all three ideas and adapts them to a Python backend so the environment can be packaged as a Docker container and hosted on Hugging Face Spaces.
+
+## Key differences from the original
+
+| | gepa-tuned-sql-agent | sql-agent-openenv (this repo) |
+|---|---|---|
+| Backend | Next.js API routes (TypeScript) | FastAPI (Python) |
+| Frontend | Next.js pages | React + Vite (static, served by FastAPI) |
+| LLM | Azure OpenAI | HF Router (Qwen 2.5-72B) |
+| Deployment | Vercel / local | Hugging Face Spaces (Docker) |
+| DB support | SQLite, PostgreSQL, MySQL | SQLite file + PostgreSQL DSN |
 
 ## OpenEnv API
 
@@ -31,6 +54,15 @@ An OpenEnv-compliant RL environment for SQL generation, featuring:
 | `/step` | POST | Execute one repair action |
 | `/state` | GET | Current environment state |
 
+## Connect your own database
+
+In the UI click **Connect DB** (top-right) and enter:
+
+- **SQLite:** `/path/to/your/database.db` or `:memory:`
+- **PostgreSQL:** `postgresql://user:password@host:5432/dbname`
+
+The agent auto-detects the dialect and adjusts its prompt accordingly.
+
 ## Demo
 
 Click **Demo** in the top-right to watch the agent fail, self-repair via RL, then improve through two GEPA prompt-evolution cycles (42% → 91%).