Natural-Language-to-SQL API

Turn questions into queries

Most people in a company can describe what they want to know but can't write the SQL to get it. unoblox turns a plain-English question into a query against your schema — through one OpenAI-compatible endpoint, billed in rupees.

Why text-to-SQL

SQL is powerful but gatekept by syntax. A language model that's shown your schema can translate "how many customers signed up last month" into a working query, putting the database in reach of anyone who can type a question.

Best models

Model	Input ₹/1M	Output ₹/1M	Fit
Qwen3 Max	₹120.95	₹604.77	Cost-effective for simple, repetitive queries
GPT-4.1	₹201.6	₹806.4	Strong general SQL generation
Claude Opus	₹504	₹2520	Best for multi-table joins and complex logic

Prices per 1M tokens.

Implementation

from openai import OpenAI

client = OpenAI(
    api_key="ub-gw-...",
    base_url="https://api.unoblox.ai/v1"
)

schema = (
    "TABLE users (id INT, name VARCHAR, email VARCHAR, created_at DATE)\n"
    "TABLE orders (id INT, user_id INT, total DECIMAL, status VARCHAR, created_at DATE)"
)

question = "How many customers signed up last month and how much did they spend?"

prompt = "Schema:\n" + schema + "\n\nQuestion: " + question + "\n\nReturn only a PostgreSQL query, no explanation."

response = client.chat.completions.create(
    model="openai/gpt-4-1",
    messages=[{"role": "user", "content": prompt}]
)

query = response.choices[0].message.content.strip()

Keeping it safe

Never execute a model's SQL output directly against a production database. Put guardrails around it:

1. Read-only by default: connect with a database user that can only SELECT.
2. Instruct explicitly: add "generate only SELECT statements; never DELETE, DROP, or UPDATE" to the system prompt.
3. Parse before executing: check the generated query only touches expected tables before running it.
4. Row and time limits: cap result size and query timeout so a bad join can't lock up your database.

Example questions it handles well

Question type	Example
Aggregation	"Total revenue by month this year"
Filtering	"Customers who haven't ordered in 90 days"
Joins	"Top 10 products by revenue, with category names"
Comparison	"This month's signups vs. last month's"

Use cases

• Business intelligence dashboards, no-code querying for non-engineers
• Customer support tooling, quick account and data lookups
• Internal analytics without waiting on a data team
• Self-service reporting for sales and finance

Frequently asked questions

Q: How reliable is the generated SQL? A: Simple aggregations and filters are usually correct on the first try; complex multi-table joins benefit from a stronger model (Claude Opus) and should get a human glance before running against production data.

Q: Can I use it across MySQL, PostgreSQL, or Snowflake? A: Yes — state your SQL dialect in the schema description. Most models adapt syntax correctly once told which dialect to target.

Q: What about SQL injection risk? A: The model isn't taking untrusted user input directly into a query string — it's generating the query itself. Still, always run it through a read-only, permission-limited database user, per the guardrails above.

Q: Can I restrict it from ever writing data? A: Yes — instruct it explicitly to generate only SELECT statements, and enforce that at the database permission level too, not just in the prompt.

Q: How do I handle ambiguous questions like "last month"? A: Give the model today's date and a definition ("last month" = the previous calendar month) in the system prompt, and ask it to state its assumption alongside the query.

Q: Can I ask questions in Hindi and get SQL back? A: Yes — Claude Opus and GPT-4.1 both understand Hindi and regional-language questions and will still return standard SQL.

Get started in rupees → https://unoblox.ai/sign-in