Picture by Writer# Introduction
Knowledge has turn out to be an indispensable useful resource for any profitable enterprise, because it gives helpful insights for knowledgeable decision-making. Given the significance of knowledge, many firms are constructing methods to retailer and analyze it. Nonetheless, there are a lot of instances when it’s laborious to amass and analyze the mandatory knowledge, particularly with the growing complexity of the info system.
With the appearance of generative AI, knowledge work has turn out to be considerably simpler, as we will now use easy pure language to obtain principally correct output that carefully follows the enter we offer. It’s additionally relevant to knowledge processing and evaluation with SQL, the place we will ask for question growth.
On this article, we’ll develop a easy API utility that interprets pure language into SQL queries that our database understands. We are going to use three essential instruments: OpenAI, FastAPI, and SQLite.
Right here’s the plan.
# Textual content-to-SQL App Growth
First, we’ll put together every little thing wanted for our venture. All that you must present is the OpenAI API key, which we’ll use to entry the generative mannequin. To containerize the appliance, we’ll use Docker, which you’ll purchase for the native implementation utilizing Docker Desktop.
Different parts, reminiscent of SQLite, will already be out there while you set up Python, and FastAPI might be put in later.
For the general venture construction, we’ll use the next:
text_to_sql_app/
├── app/
│ ├── __init__.py
│ ├── database.py
│ ├── openai_utils.py
│ └── essential.py
├── demo.db
├── init_db.sql
├── necessities.txt
├── Dockerfile
├── docker-compose.yml
├── .env
Create the construction like above, or you should utilize the next repository to make issues simpler. We are going to nonetheless undergo every file to achieve an understanding of the best way to develop the appliance.
Let’s begin by populating the .env file with the OpenAI API key we beforehand acquired. You are able to do that with the next code:
OPENAI_API_KEY=YOUR-API-KEY
Then, go to the necessities.txt to fill within the vital libraries we’ll use for
fastapi
uvicorn
sqlalchemy
openai
pydantic
python-dotenv
Subsequent, we transfer on to the __init__.py file, and we’ll put the next code inside:
from pathlib import Path
from dotenv import load_dotenv
load_dotenv(dotenv_path=Path(__file__).resolve().mother or father.mother or father / ".env", override=False)
The code above ensures that the surroundings incorporates all the mandatory keys we’d like.
Then, we’ll develop Python code within the database.py file to hook up with the SQLite database we’ll create later (known asdemo.db) and supply a option to run SQL queries.
from sqlalchemy import create_engine, textual content
from sqlalchemy.orm import Session
ENGINE = create_engine("sqlite:///demo.db", future=True, echo=False)
def run_query(sql: str) -> checklist[dict]:
with Session(ENGINE) as session:
rows = session.execute(textual content(sql)).mappings().all()
return [dict(r) for r in rows]
After that, we’ll put together the openai_utils.py file that may settle for the database schema and the enter questions. The output might be JSON containing the SQL question (with a guard to forestall any write operations).
import os
import json
from openai import OpenAI
shopper = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
_SYSTEM_PROMPT = """
You change natural-language questions into read-only SQLite SQL.
By no means output INSERT / UPDATE / DELETE.
Return JSON: { "sql": "..." }.
"""
def text_to_sql(query: str, schema: str) -> str:
response = shopper.chat.completions.create(
mannequin="gpt-4o-mini",
temperature=0.1,
response_format={"sort": "json_object"},
messages=[
{"role": "system", "content": _SYSTEM_PROMPT},
{"role": "user",
"content": f"schema:n{schema}nnquestion: {question}"}
]
)
payload = json.hundreds(response.decisions[0].message.content material)
return payload["sql"]
With each the code and the connection prepared, we’ll put together the appliance utilizing FastAPI. The applying will settle for pure language questions and the database schema, convert them into SQL SELECT queries, run them by means of the SQLite database, and return the outcomes as JSON. The applying might be an API we will entry by way of the CLI.
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from sqlalchemy import examine
from .database import ENGINE, run_query
from .openai_utils import text_to_sql
app = FastAPI(title="Textual content-to-SQL Demo")
class NLRequest(BaseModel):
query: str
@app.on_event("startup")
def capture_schema() -> None:
insp = examine(ENGINE)
international SCHEMA_STR
SCHEMA_STR = "n".be part of(
f"CREATE TABLE {t} ({', '.be part of(c['name'] for c in insp.get_columns(t))});"
for t in insp.get_table_names()
)
@app.put up("/question")
def question(req: NLRequest):
strive:
sql = text_to_sql(req.query, SCHEMA_STR)
if not sql.lstrip().decrease().startswith("choose"):
increase ValueError("Solely SELECT statements are allowed")
return {"sql": sql, "outcome": run_query(sql)}
besides Exception as e:
increase HTTPException(status_code=400, element=str(e))
That’s every little thing we’d like for the primary utility. The subsequent factor we’ll put together is the database. Use the database beneath within the init_db.sql for instance functions, however you may at all times change it if you need.
DROP TABLE IF EXISTS order_items;
DROP TABLE IF EXISTS orders;
DROP TABLE IF EXISTS funds;
DROP TABLE IF EXISTS merchandise;
DROP TABLE IF EXISTS clients;
CREATE TABLE clients (
id INTEGER PRIMARY KEY,
identify TEXT NOT NULL,
nation TEXT,
signup_date DATE
);
CREATE TABLE merchandise (
id INTEGER PRIMARY KEY,
identify TEXT NOT NULL,
class TEXT,
worth REAL
);
CREATE TABLE orders (
id INTEGER PRIMARY KEY,
customer_id INTEGER,
order_date DATE,
complete REAL,
FOREIGN KEY (customer_id) REFERENCES clients(id)
);
CREATE TABLE order_items (
order_id INTEGER,
product_id INTEGER,
amount INTEGER,
unit_price REAL,
PRIMARY KEY (order_id, product_id),
FOREIGN KEY (order_id) REFERENCES orders(id),
FOREIGN KEY (product_id) REFERENCES merchandise(id)
);
CREATE TABLE funds (
id INTEGER PRIMARY KEY,
order_id INTEGER,
payment_date DATE,
quantity REAL,
technique TEXT,
FOREIGN KEY (order_id) REFERENCES orders(id)
);
INSERT INTO clients (id, identify, nation, signup_date) VALUES
(1,'Alice','USA','2024-01-05'),
(2,'Bob','UK','2024-03-10'),
(3,'Choi','KR','2024-06-22'),
(4,'Dara','ID','2025-01-15');
INSERT INTO merchandise (id, identify, class, worth) VALUES
(1,'Laptop computer Professional','Electronics',1500.00),
(2,'Noise-Canceling Headphones','Electronics',300.00),
(3,'Standing Desk','Furnishings',450.00),
(4,'Ergonomic Chair','Furnishings',250.00),
(5,'Monitor 27"','Electronics',350.00);
INSERT INTO orders (id, customer_id, order_date, complete) VALUES
(1,1,'2025-02-01',1850.00),
(2,2,'2025-02-03',600.00),
(3,3,'2025-02-05',350.00),
(4,1,'2025-02-07',450.00);
INSERT INTO order_items (order_id, product_id, amount, unit_price) VALUES
(1,1,1,1500.00),
(1,2,1,300.00),
(1,5,1,350.00),
(2,3,1,450.00),
(2,4,1,250.00),
(3,5,1,350.00),
(4,3,1,450.00);
INSERT INTO funds (id, order_id, payment_date, quantity, technique) VALUES
(1,1,'2025-02-01',1850.00,'Credit score Card'),
(2,2,'2025-02-03',600.00,'PayPal'),
(3,3,'2025-02-05',350.00,'Credit score Card'),
(4,4,'2025-02-07',450.00,'Financial institution Switch');
Then, run the next code in your CLI to create a SQLite database for our venture.
With the database prepared, we’ll create a Dockerfile to containerize our utility.
FROM python:3.12-slim
WORKDIR /code
COPY necessities.txt .
RUN pip set up --no-cache-dir -r necessities.txt
COPY . .
CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "8000"]
We can even create a docker-compose.yml file for operating the appliance extra easily.
providers:
text2sql:
construct: .
env_file: .env
ports:
- "8000:8000"
restart: unless-stopped
volumes:
- ./demo.db:/code/demo.db
With every little thing prepared, begin your Docker Desktop and run the next code to construct the appliance.
docker compose construct --no-cache
docker compose up -d
If every little thing is completed nicely, you may take a look at the appliance by utilizing the next code. We are going to ask what number of clients now we have within the knowledge.
curl -X POST "http://localhost:8000/question" -H "Content material-Sort: utility/json" -d "{"query":"What number of clients?"}"
The output will appear to be this.
{"sql":"SELECT COUNT(*) AS customer_count FROM clients;","outcome":[{"customer_count":4}]}
We are able to strive one thing extra advanced, just like the variety of orders for every buyer:
curl -X POST "http://localhost:8000/question" -H "Content material-Sort: utility/json" -d "{"query":"What's the variety of orders positioned by every buyer"}"
With output like beneath.
{"sql":"SELECT customer_id, COUNT(*) AS number_of_orders FROM orders GROUP BY customer_id;","outcome":[{"customer_id":1,"number_of_orders":2},{"customer_id":2,"number_of_orders":1},{"customer_id":3,"number_of_orders":1}]}
That’s all that you must construct a primary Textual content-to-SQL utility. You’ll be able to improve it additional with a front-end interface and a extra advanced system tailor-made to your wants.
# Wrapping Up
Knowledge is the center of any knowledge work, and corporations use it to make choices. Many instances, the system now we have is just too advanced, and we have to depend on generative AI to assist us navigate it.
On this article, now we have realized the best way to develop a easy Textual content-to-SQL utility utilizing the OpenAI mannequin, FastAPI, and SQLite.
I hope this has helped!
Cornellius Yudha Wijaya is an information science assistant supervisor and knowledge author. Whereas working full-time at Allianz Indonesia, he likes to share Python and knowledge ideas by way of social media and writing media. Cornellius writes on quite a lot of AI and machine studying subjects.
