About SiyasatQL

Ask questions about a real e-commerce database in plain English. SiyasatQL writes a read-only SQL query, shows you its steps, and waits for you to review it and click Run. Then you get a table and, when one fits, a chart.

The data comes from Olist, a Brazilian online marketplace where small sellers list products. Each order records what was bought, from which seller, how it was paid and shipped, and what the customer thought of it.

orders

99,441

unique customers

96,096

sellers

3,095

products

32,951

• customers
• orders
• order_items
• order_payments
• order_reviews
• products
• sellers
• product_category_name_translation

Covers September 2016 to October 2018 (fullest from January 2017 to August 2018). Money is in Brazilian reais (BRL). Each question stands alone: SiyasatQL doesn't remember earlier ones.

How accurate is it?

It was tested on 60 questions with known answers, from simple counts to multi-step ones. Over three runs, the setup used here (Claude Haiku 5.5) returned the correct result 96.7% of the time on average (95.0% to 98.3%). It got every easy question right; medium questions were its weakest tier at 93.1%.

It was also tried with 14 unsafe prompts, such as requests to change data or hide commands inside a question. None produced SQL that would change anything, and the database refuses writes on its own anyway.

These numbers describe this 60-question set on this dataset, not text-to-SQL in general, and the model varies from run to run, so an occasional wrong answer is expected. Always read the SQL before you run it. Full evaluation details.

SiyasatQL was built by Romulo Ramos Jr. To get in touch, email romu.dev@outlook.com or send a message on LinkedIn.