txt2sql.py

import sqlite3
from sqlite3 import Error
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer, BitsAndBytesConfig
from transformers import AutoModelForCausalLM
from openai import OpenAI
import google.generativeai as genai

class SQLPromptModel:
    def __init__(self, model_dir, database):
        self.model_dir = model_dir
        self.database = database
        # peft_model_dir = self.model_dir
        bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype="float16",
            bnb_4bit_use_double_quant=True,
        )
        # self.model = AutoPeftModelForCausalLM.from_pretrained(
        #     peft_model_dir, low_cpu_mem_usage=True, quantization_config=bnb_config
        # )
        # self.tokenizer = AutoTokenizer.from_pretrained(peft_model_dir)
        # self.model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
        # self.tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
        self.chatgpt_client = OpenAI(api_key="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")
        self.genai = genai
        self.genai.configure(api_key="AIzaSyAFG94rVbm9eWepO5uPGsMha8XJ-sHbMdA")
        self.genai_model = genai.GenerativeModel('gemini-pro')

        self.conn = sqlite3.connect(self.database)

    def fetch_table_schema(self, table_name):
        """Fetch the schema of a table from the database."""
        cursor = self.conn.cursor()
        cursor.execute(f"PRAGMA table_info({table_name})")
        schema = cursor.fetchall()
        if schema:
            return schema
        else:
            print(f"Table {table_name} does not exist or has no schema.")
            return None
        
    def text2sql(self, schema, user_prompt, inp_prompt=None):
        """Generate SQL query based on user prompt and table schema.inp_prompt is for gradio purpose"""
        table_columns = ', '.join([f"{col[1]} {col[2]}" for col in schema])

        prompt = f"""Below are SQL table schemas paired with instructions that describe a task.
        Using valid SQLite, write a response that appropriately completes the request for the provided tables.
        Select all columns unless specified in specific.
        Example row :1,Michael,michael@ignitarium.com,59,Female,Headphones,2023-01-03,General inquiry,Email,44 hours,88 hours,4,Technical Issue,Server crashes due to memory leaks in custom-developed software.,Closed,"Debug and optimize the software code to identify and fix memory leaks, and implement regular monitoring for early detection.",Medium
        ### Instruction: {user_prompt} ### 
        Input: CREATE TABLE ticket_dataset({table_columns});
        ### Response: (Return only query , nothing extra)"""

        if inp_prompt is not None :
            prompt = prompt.replace(user_prompt, inp_prompt + " ")
        else:
            inp_prompt = input("Press Enter for default question or Enter user prompt without newline characters: ").strip()
            if inp_prompt:
                prompt = prompt.replace(user_prompt, inp_prompt + " ")

        """Text to SQL query generation"""
        input_ids = self.tokenizer(
            prompt, return_tensors="pt", truncation=True
        ).input_ids.to(next(self.model.parameters()).device)  # Move input to the device of the model
        outputs = self.model.generate(input_ids=input_ids, max_new_tokens=200)
        response = self.tokenizer.batch_decode(
            outputs.detach().cpu().numpy(), skip_special_tokens=True
        )[0][:]
        return response[len(prompt):]

    def text2sql_chatgpt(self, schema, user_prompt, inp_prompt=None):
        table_columns = ', '.join([f"{col[1]} {col[2]}" for col in schema])

        prompt = f"""Below are SQL table schemas paired with instructions that describe a task.
        Using valid SQLite, write a response that appropriately completes the request for the provided tables.
        Select all columns unless specified in specific.
        Example row :1,Michael,michael@ignitarium.com,59,Female,Headphones,2023-01-03,General inquiry,Email,44 hours,88 hours,4,Technical Issue,Server crashes due to memory leaks in custom-developed software.,Closed,"Debug and optimize the software code to identify and fix memory leaks, and implement regular monitoring for early detection.",Medium
        ### Instruction: {user_prompt} ### 
        Input: CREATE TABLE ticket_dataset({table_columns});
        ### Response: (Return only generated query based on user_prompt , nothing extra)"""

        if inp_prompt is not None :
            prompt = prompt.replace(user_prompt, inp_prompt + " ")
        else:
            inp_prompt = input("Press Enter for default question or Enter user prompt without newline characters: ").strip()
            if inp_prompt:
                prompt = prompt.replace(user_prompt, inp_prompt + " ")
        print(prompt)
        completion = self.chatgpt_client.chat.completions.create(
            model="gpt-3.5-turbo",
            messages=[
                {"role": "system", "content": "You are a expert SQL developer , generate a sql query and return it"},
                {"role": "user", "content": prompt }
            ]
        )
        return completion.choices[0].message.content

    def text2sql_gemini(self, schema, user_prompt, inp_prompt=None):
        table_columns = ', '.join([f"{col[1]} {col[2]}" for col in schema])

        prompt = f"""Below are SQL table schemas paired with instructions that describe a task.
        Using valid SQLite, write a response that appropriately completes the request for the provided tables.
        Select all columns unless specified in specific.
        Example row :1,Michael,michael@ignitarium.com,59,Female,Headphones,2023-01-03,General inquiry,Email,44 hours,88 hours,4,Technical Issue,Server crashes due to memory leaks in custom-developed software.,Closed,"Debug and optimize the software code to identify and fix memory leaks, and implement regular monitoring for early detection.",Medium
        ### Instruction: {user_prompt} ### 
        Input: CREATE TABLE ticket_dataset({table_columns});
        ### Response: (Return only generated query based on user_prompt , nothing extra)"""

        if inp_prompt is not None :
            prompt = prompt.replace(user_prompt, inp_prompt + " ")
        else:
            inp_prompt = input("Press Enter for default question or Enter user prompt without newline characters: ").strip()
            if inp_prompt:
                prompt = prompt.replace(user_prompt, inp_prompt + " ")
        print(prompt)
        completion = self.genai_model.generate_content(prompt)
        generated_query=completion.text
        start_index = generated_query.find("SELECT")
        end_index = generated_query.find(";", start_index) + 1
        print(start_index,end_index)
        if start_index != -1 and end_index != 0:
            return generated_query[start_index:end_index]
        else:
            return generated_query
        


    def execute_query(self, query):
        """Executing the query on database and returning rows and columns."""
        print(query)
        cur = self.conn.cursor()
        cur.execute(query)
        col = [header[0] for header in cur.description]
        dash = "-" * sum(len(col_name) + 4 for col_name in col)
        print(tuple(col))
        print(dash)
        rows = []
        for member in cur:
            rows.append(member)
            print(member)
        cur.close()
        self.conn.commit()
        # print(rows)
        return rows, col

if __name__ == "__main__":
    model_dir = "multi_table_demo/checkpoint-2600"
    database = r"ticket_dataset.db"
    sql_model = SQLPromptModel(model_dir, database)
    user_prompt = "Give complete details of properties in India"
    while True:
        table_schema = sql_model.fetch_table_schema("ticket_dataset")
        if table_schema:
            # query = sql_model.text2sql(table_schema, user_prompt)
            # query = sql_model.text2sql_chatgpt(table_schema, user_prompt)
            query = sql_model.text2sql_gemini(table_schema, user_prompt)
            print(query)
            sql_model.execute_query(query)
            
    sql_model.conn.close()