nonoql

NoNoQL - Natural Language to SQL/MongoDB Query Generator

NoNoQL (formerly TexQL) is a T5-based transformer model that converts natural language queries into both SQL and MongoDB queries. It supports SELECT, INSERT, UPDATE, DELETE, and other database operations.

🎯 Model Description

This model translates natural language database queries into syntactically correct SQL and MongoDB commands. It's trained on a custom dataset of 30,000+ query pairs covering various database operations, tables, and query patterns.

Key Features

• ✅ Dual Output: Generates both SQL and MongoDB queries from a single natural language input
• ✅ Multi-Operation Support: SELECT, INSERT, UPDATE, DELETE, CREATE TABLE, and more
• ✅ Comparison Operators: Handles greater than, less than, equal to, and other comparisons
• ✅ Complex Queries: Supports WHERE clauses, aggregations, ordering, and limiting
• ✅ Post-Processing: Includes fixes for common model hallucinations and syntax errors

📊 Model Details

• Model Architecture: T5 (Text-to-Text Transfer Transformer)
• Base Model: google/t5-small
• Parameters: ~60M
• Training Data: 30,000+ natural language to SQL/MongoDB query pairs
• Training Strategy: Unified model trained on both SQL and MongoDB simultaneously
• Input Format: translate to {sql|mongodb}: {natural_language_query}

Supported Tables/Collections

• employees: employee_id, name, email, department, salary, hire_date, age
• departments: department_id, department_name, manager_id, budget, location
• projects: project_id, project_name, start_date, end_date, budget, status
• orders: order_id, customer_name, product_name, quantity, order_date, total_amount
• products: product_id, product_name, category, price, stock_quantity, supplier

🚀 Usage

Installation

pip install transformers torch

Basic Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Load model and tokenizer
model_name = "mohhhhhit/nonoql"  # Replace with your HF model path
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

# Generate SQL query
def generate_query(natural_language, target_type='sql'):
    input_text = f"translate to {target_type}: {natural_language}"
    inputs = tokenizer(input_text, return_tensors="pt", max_length=256, truncation=True)
    
    outputs = model.generate(
        **inputs,
        max_length=512,
        num_beams=10,
        temperature=0.3,
        repetition_penalty=1.2,
        length_penalty=0.8,
        early_stopping=True
    )
    
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Example usage
nl_query = "Find employees where salary is greater than 50000"

sql_query = generate_query(nl_query, target_type='sql')
print(f"SQL: {sql_query}")
# Output: SELECT * FROM employees WHERE salary > 50000;

mongodb_query = generate_query(nl_query, target_type='mongodb')
print(f"MongoDB: {mongodb_query}")
# Output: db.employees.find({"salary": {$gt: 50000}});

Example Queries

🎓 Training

Dataset

• Size: 30,000+ query pairs
• Operations: SELECT (40%), INSERT (20%), UPDATE (20%), DELETE (15%), CREATE (5%)
• Tables: 5 main tables with realistic schemas
• Generation: Synthetic data with varied patterns and complexity

Training Configuration

training_args = {
    "learning_rate": 3e-4,
    "per_device_train_batch_size": 8,
    "per_device_eval_batch_size": 8,
    "num_train_epochs": 10,
    "weight_decay": 0.01,
    "warmup_steps": 500,
    "max_seq_length": 512,
}

Evaluation Metrics

• BLEU Score: ~85%
• Exact Match: ~78%
• Syntax Correctness: ~92% (after post-processing)

⚙️ Post-Processing

The model includes several post-processing fixes to handle common issues:

1. Comparison Operators: Converts = to >, <, >=, <= based on keywords like "greater than", "less than"
2. Operation Type: Fixes wrong operations (e.g., SELECT when DELETE is intended)
3. MongoDB Syntax: Adds missing curly braces and converts to proper MongoDB operators
4. UPDATE Queries: Reconstructs malformed UPDATE statements
5. CREATE TABLE: Fixes hallucinated columns in table creation

⚠️ Limitations

• Schema Awareness: Model is trained on specific tables; may not generalize to completely new schemas
• Complex Joins: Limited support for multi-table JOINs and subqueries
• Advanced Features: May struggle with window functions, CTEs, and advanced SQL features
• Hallucinations: Can generate incorrect column names for unseen patterns (mitigated by post-processing)
• Case Sensitivity: Works best with lowercase natural language inputs

📝 Known Issues & Fixes

🛠️ Use Cases

• Database Query Assistants: Help non-technical users query databases
• Educational Tools: Teach SQL/MongoDB syntax through examples
• Prototyping: Quickly generate queries for testing
• Documentation: Auto-generate query examples
• Migration Tools: Convert between SQL and MongoDB syntaxes

📄 Citation

If you use this model in your research or application, please cite:

@misc{nonoql2026,
  title={NoNoQL: Natural Language to SQL and MongoDB Query Generation},
  author={Mohit Panchal},
  year={2026},
  howpublished={\url{https://huggingface.co/mohhhhhit/nonoql}},
}

📜 License

This model is released under the Apache 2.0 License.

🤝 Contributing

Contributions, feedback, and suggestions are welcome! Please feel free to:

• Report issues or bugs
• Suggest new features
• Improve the training data
• Add support for more database systems

🔗 Links

• Model Repository: Hugging Face
• GitHub: Source Code
• Demo: Streamlit App

🙏 Acknowledgments

• Built on the T5 architecture by Google Research
• Trained using the Hugging Face Transformers library
• Inspired by the need for more accessible database querying tools

---

Note: This model is designed for educational and prototyping purposes. Always validate generated queries before executing them on production databases.