Text2Sql using RAG & Finetuning microsoft/phi-2

Overview

1. LoRA-based Fine-tuning for optimizing language models.
2. RAG (Retrieval-Augmented Generation) System for efficient text-to-SQL conversion using vector stores and NeonDB.

---

1. Fine-tuning Module

This module utilizes LoRA (Low-Rank Adaptation) to fine-tune the Phi-2 model efficiently. Key features include:

Features

• Uses LoRAConfig for lightweight fine-tuning with optimal performance.
• Fine-tunes the microsoft/phi-2 model with quantization enabled for CPU efficiency.
• LoRA configuration settings ensure control over trainable parameters, scaling factor, and dropout rate for effective adaptation.

---

2. RAG (Retrieval-Augmented Generation) System

This module extracts text from document files to generate SQL queries using vector stores and a conversational AI system.

Features

• Supports .pdf, .docx, and .txt document types for schema extraction.
• Utilizes PGVector for efficient vector storage and retrieval of schema details and sample queries.
• Uses the Groq API to generate SQL queries based on provided schemas and example queries.
• Implements retry logic to handle API failures and ensure stability.

Key Libraries Used

• gradio (for UI)
• pdfplumber, docx, json (for text extraction)
• langchain_postgres (for PGVector storage)

---

Installation

1. Clone the repository:

git clone <repository-url>
cd finetunee

2. Install dependencies:

pip install -r requirements.txt

API Keys & Database Configuration

• Ensure your Groq API Key and NeonDB Connection String are correctly set in the environment variables for secure access.

---