Text2SQL — Group 28 (CSE 573)

A natural language to SQL system with a Next.js dashboard frontend and a Python ML pipeline backend using RAG, schema linking, and back-translation validation.

---

Repository Structure

text2sql/
├── frontend/       ← Next.js 14 dashboard (TypeScript + Tailwind)
├── backend/        ← FastAPI server
└── ml/             ← Python Text2SQL pipeline (Groq + FAISS + SQLite)

---

Frontend

cd frontend
npm install
npm run dev         # http://localhost:3000

The frontend connects to a backend API. Override the default URL by setting NEXT_PUBLIC_API_URL in frontend/.env.local.

---

Backend

cd backend
pip install -r requirements.txt
uvicorn main:app --reload  # http://localhost:8000

Evaluation

Results, diagnostics, and experiment notebooks are in evaluation/. See evaluation/EVALUATION_SUMMARY.md for full results.

ML Pipeline

1. Download data

Download the ml_data folder from Google Drive, rename it to data, and place it inside ml/:

ml/
└── data/
    ├── database/               ← Spider SQLite databases
    ├── tables.json             ← schema definitions
    ├── all_graphs.pkl          ← Steiner tree schema graphs
    ├── faiss_questions.index   ← FAISS question index
    ├── faiss_schemas.index     ← FAISS schema index
    ├── questions_metadata.pkl  ← question metadata
    └── schemas_metadata.pkl    ← schema metadata

2. Create .env

Create ml/.env with your Groq API key (free at console.groq.com):

GROQ_API_KEY=your_key_here

3. Install dependencies

cd ml
python -m venv .venv
source .venv/bin/activate        # Mac/Linux
# .venv\Scripts\activate         # Windows
pip install -r requirements.txt

4. Test the pipeline

python pipeline.py

You should see SQL queries and results printed for 3 test questions.

---

ML Usage

Basic usage

from pipeline import setup, run

ctx = setup()  # call once at startup

result = run(
    question = "How many singers are there?",
    db_id    = "concert_singer",
    ctx      = ctx,
    pipeline = "rag_bt"   # "raw" | "rag" | "rag_bt"
)

print(result["sql"])     # generated SQL
print(result["result"])  # list of rows from DB
print(result["error"])   # None if success

Available pipelines

Pipeline	Description
raw	Zero-shot — schema + question only, no examples
rag	RAG + 3-level repair (syntax, schema, execution)
rag_bt	RAG + repair + back-translation semantic validation

Result dictionary

Key	Type	Description
sql	str	Generated SQL query
result	list	Rows returned from database execution
error	str/None	Execution error if any, else None
pipeline	str	Which pipeline was used
repairs	int	Number of repair attempts made
bt_match	bool/None	Back-translation match result (rag_bt only)
issue	str/None	Semantic issue detected (rag_bt only)

Get available databases

from pipeline import setup, get_available_databases

ctx = setup()
dbs = get_available_databases(ctx)
print(dbs)  # ['academic', 'airline', 'car_1', 'concert_singer', ...]

---

FastAPI Integration

import sys
sys.path.append("../ml")

from fastapi import FastAPI
from pipeline import setup, run, get_available_databases

app = FastAPI()
ctx = setup()  # loads data once on startup

@app.get("/databases")
def databases():
    return {"databases": get_available_databases(ctx)}

@app.post("/query")
def query(question: str, db_id: str, pipeline: str = "rag_bt"):
    return run(question, db_id, ctx, pipeline=pipeline)

---

Pipeline Architecture

User Question
      │
      ▼
Schema Linking (deterministic keyword match → terminal tables)
      │
      ▼
RAG Retrieval (FAISS → top-3 similar examples)
      │
      ▼
Steiner Tree (schema-grounded JOIN path for multi-table queries)
      │
      ▼
LLM Generation
      │
      ▼
Level 1: Syntax Check
Level 2: Schema Validation
Level 3: Execution Check
      │ (repair loop up to 3 attempts if any level fails)
      ▼
Level 4: Back-Translation Validation (rag_bt only)
   LLM describes SQL in English → checks against original question
   If mismatch → targeted semantic repair
      │
      ▼
Final SQL + Result

Uses Llama 3.3 70B via Groq API (free tier). No local model download required.

---

Notes

• data/ and .env are not committed — download data from Google Drive and create .env manually
• setup() takes ~10 seconds on first run (loads FAISS indexes and graphs)
• Each run() call makes 2–4 API calls to Groq depending on pipeline and repair needs