The recent advancements in natural language processing have led to the emergence of Large Language Models (LLMs), which have shown impressive performance in tasks like text generation, translation, and several other tasks. These advancements have also extended to database interactions, which enable users to interact with SQL databases using simple, everyday language. Significant attention has been drawn to these language models across different fields due to their capability to produce text like humans. The emergence of transformer-based structures, as showcased by the GPT-3.5 model, has paved the way for a new age in language understanding, exceeding previous constraints and enabling a deeper understanding of context.

Large Language Models (LLMs) are a revolutionary advancement that links human language and SQL, and may, in the future, form the interface to SQL databases for humans. Through the interpretation of queries in natural language, LLM can create SQL commands. This makes it a great choice for users with minimal database knowledge.

Advantages of LLMs in SQL Data Interactions

LLMs make database interactions accessible for users with non-technical skills by simplifying SQL data tasks.
LLMs let users communicate with databases in their language, removing the need for exact command syntax.
By leveraging their superior understanding of language abilities, LLMs can extract contextual data from user queries, which results in the generation of more precise SQL queries.
Mastering SQL requires heavy training, and LLMs provide a more intuitive alternative by eliminating the learning process and accelerating data access and updates.

As we progress through the sections, we will examine the practical elements of establishing the environment, performing queries, updates, inserts, and deleting SQL data with LLMs. We will also analyze optimal methods and factors to consider while using LLMs for SQL database interaction. Upon completion of this article, readers will be able to comprehend and efficiently use LLMs in their database dealings, which can transform their interaction with SQL data.

Setting Up the Environment

There needs to be a conducive environment to begin interacting with the SQL database using LLMs. This section outlines a step-by-step process for setting up the environment to enable interaction between your PostgreSQL database and DB-GPT pre-trained LLM model.

Head over to MyAccount portal on E2E Cloud.
Install a GPU node. It comes pre-installed with the right drivers.
Make sure Python is installed (preferably a 3.6 version or higher) on your machine.
Install the transformer-based library with `pip install transformers`.
Install the ‘psycopg2’ library that allows communication with the PostgreSQL database using `pip install psycopg2`.
You need to retrieve the necessary credentials for your system, like hostname, port, database name, username, and password.
To establish a connection to your PostgreSQL database, use the psycopg2 library as shown below:


import psycopg2 
host = 'your_postgresql_host' 
port = 'your_postgresql_port' 
database = 'your_postgresql_database' 
username = 'your_postgresql_username' 
password = 'your_postgresql_password' 
connection_string = f'host={host} port={port} dbname={database} user={username} password={password}' 
connection = psycopg2.connect(connection_string)

Load the DB-GBT pre-trained model using the transformers library


from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model_name = 'eosphoros/db-gpt'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

The environment is now prepared to mediate interactions with the PostgreSQL database using the ‘DB-GPT’ pre-trained LLM model. The following sections will show how to use this setup for natural language querying, updating, inserting, and deleting data within the PostgreSQL database using LLMs.

Generating and Executing SQL Queries

The use of Language Models and the 'DB-GPT' pre-trained Language Model for Databases for engaging with SQL data provides an easy-to-use method for database queries. In this part, we'll show you how to use an LLM to perform SQL queries from natural language inputs. An LLM is engineered to understand and interpret natural language queries, enabling users to articulate their data retrieval requirements in a simple language instead of depending on intricate SQL syntax. The model's advanced language comprehension abilities allow it to formulate equivalent SQL queries that accurately extract the necessary information from the database. For creating SQL queries utilizing LLM, follow these instructions below:

‍

Request the user to formulate their query in a natural language format.
Use the LLM's tokenizer to tokenize the input provided by the user, making it ready for model input.
Channel the tokenized input through the LLM model to produce the SQL query.
Convert the SQL query generated by the model from its tokenized form into a SQL syntax that is understandable to humans.

‍

The SQL query obtained from an LLM needs to be run against the linked PostgreSQL database to fetch the results. The following code snippet provides an example of the process:


def execute_sql_query(connection, query):
    try:
        cursor = connection.cursor()
        cursor.execute(query)
        result = cursor.fetchall()
        cursor.close()
        return result
    except (Exception, psycopg2.Error) as error:
        print('Error executing the query:', error)
        return None

user_query = 'Insert query here.'

# Tokenize and generate SQL query
encoded_input = tokenizer(user_query, return_tensors='pt')
with torch.no_grad():
    output = model.generate(**encoded_input)
decoded_query = tokenizer.decode(output[0], skip_special_tokens=True)

result_set = execute_sql_query(connection, decoded_query)

# Display the results
if result_set:
    for row in result_set:
        print(row)
else:
    print('No results found.')

Utilizing LLM for SQL data queries allows users to effectively retrieve information from the database without mastering complex SQL syntax. This makes the data retrieval procedure more user-friendly and intuitive.

Executing Updated SQL Statements on the Database

To update your database, follow these steps:

Request the user to provide the details they wish to implement and the criteria that pinpoint the records for updating.
Utilize LLM's tokenizer to break down the user's input and ready it for model input.
Feed the tokenized input into the LLM model to create the SQL update statement.
Convert the SQL update statement borne by the model from its tokenized form into an understandable SQL syntax.

After successfully getting the updated statement, refer to the following code snippet to apply the modifications:


def execute_sql_update(connection, update_statement):
    try:
        cursor = connection.cursor()
        cursor.execute(update_statement)
        connection.commit()
        cursor.close()
        print('Update successful!')
    except (Exception, psycopg2.Error) as error:
        connection.rollback()
        print('Error executing the update:', error)

# Updated instructions
user_update_instruction = 'Write your updated instruction here.'

encoded_input = tokenizer(user_update_instruction, return_tensors='pt')
with torch.no_grad():
    output = model.generate(**encoded_input)
decoded_update_statement = tokenizer.decode(output[0], skip_special_tokens=True)

execute_sql_update(connection, decoded_update_statement)

This helps in conveniently altering the database records without needing to craft complex update queries. This strategy simplifies the data management process, allowing users to engage with the database through straightforward and intuitive language instructions.

Inserting and Deleting the SQL Data

Users can use natural language instructions to detail the data they wish to include when adding new records to the database. Since any LLM has language skills, it can translate these descriptions and produce the corresponding results to the inserted SQL statements. Follow these steps to insert the data by using LLM:

Ask the user for a description in the natural language of the data they wish to input.
Use the LLM's tokenizer to break down the user's input into tokens for model processing.
Feed the tokenized input into the LLM model to create the SQL insert statement.
Convert the SQL insert statement generated by the model from its tokenized form to a human-friendly SQL syntax.

After obtaining the insert statement, execute the following:


def execute_sql_insert(connection, insert_statement):
    try:
        cursor = connection.cursor()
        cursor.execute(insert_statement)
        connection.commit()
        cursor.close()
        print('Insertion successful!')
    except (Exception, psycopg2.Error) as error:
        connection.rollback()
        print('Error executing the insert:', error)

# User's natural language description of data to insert
user_insert_description = 'Insert new natural language description for updating here.'

encoded_input = tokenizer(user_insert_description, return_tensors='pt')
with torch.no_grad():
    output = model.generate(**encoded_input)
decoded_insert_statement = tokenizer.decode(output[0], skip_special_tokens=True)

execute_sql_insert(connection, decoded_insert_statement)

Similarly, use natural language instructions to detail the data that needs to be deleted and the LLM can convert these descriptions to produce the required results. Follow the site steps as before and obtain the delete statement.

After getting the delete statement, implement the following code:


def execute_sql_delete(connection, delete_statement):
    try:
        cursor = connection.cursor()
        cursor.execute(delete_statement)
        connection.commit()
        cursor.close()
        print('Deletion successful!')
    except (Exception, psycopg2.Error) as error:
        connection.rollback()
        print('Error executing the delete:', error)

# Natural language instruction for data deletion
user_delete_instruction = 'Insert new natural language description for deleting here.'

encoded_input = tokenizer(user_delete_instruction, return_tensors='pt')
with torch.no_grad():
    output = model.generate(**encoded_input)
decoded_delete_statement = tokenizer.decode(output[0], skip_special_tokens=True)

execute_sql_delete(connection, decoded_delete_statement)

The effortless implementation of operations like data querying, updating, inserting, and deleting within the SQL database is made possible for users of various skill levels through LLM's seamless integration of natural language instructions. This article has displayed the practical use of LLMs, by making use of an open-source tech stack to maintain transparency and reproducibility. By following the instructions in this article, one can create an optimal environment for an LLM, link to their PostgreSQL database, and interact with it efficiently.

You can try this today on a node on E2E Cloud. Simply head over to MyAccount, login or register, spin up a node, and test the above in a Jupyter Notebook.

Nigama Vykari

POSTS BY THIS AUTHOR

How to Use LLMs To Interact With Your SQL Data