COMING SOON · POWERED BY JEV
Jev picks the tables.
AI2SQL writes the SQL.
Big schemas break text-to-SQL. Before any SQL is written, one Jev call reads your question and ranks every table at once, so the LLM only sees the few it actually needs.
Video transcript
Your database has 56 tables. Your question needs 3.
A prompt with the full schema: 56 table definitions. Slow. Expensive. Easy to grab the wrong table, like sandbox_orders instead of orders.
Question: What was our monthly revenue by customer country last year?
Jev, one call, 56 options. Top tables: orders 0.36, customers 0.35, payments 0.29. One call ranks all 56 tables.
The LLM writes the SQL with 3 tables, not 56: monthly revenue by customer country from payments joined to orders and customers.
Same 56-table schema, one ranked call vs. 56 yes/no calls: every required table in the top 4, vs. as deep as #46 when asked one table at a time. 1 call instead of 56. 1/10 the tokens.
Jev picks the tables. AI2SQL writes the query. Coming soon. Join the waitlist at ai2sql.io/jev. Powered by Jev (TypeSafe).
What is Jev table selection?
Jev table selection is a step AI2SQL will run before writing any SQL. Your question and your schema's table and column names go to Jev, a decision model from TypeSafe AI, in a single call. Jev returns a probability for every table, and the LLM that writes the query gets only the top few.
This matters because text-to-SQL usually fails on the schema, not the SQL. Put 56 tables in the prompt and the model gets slower, costs more, and can join sandbox_orders where it needed orders. Asking about each table separately does not fix it: a table judged on its own has nothing to be compared against.
Results on a 56-table schema
How it was tested
- An e-commerce schema grown to 56 tables with the names real databases collect, such as sandbox_orders, invoices_legacy and billing_plans_v1.
- Three questions with known answers, each run twice, in September 2026.
- One Jev question with all 56 tables as options, compared against 56 separate yes/no questions.
- The measure was the worst rank of a table the query actually needs. Method and code are in the write-up.
How it works
You ask a question
In plain English, against a schema with dozens or hundreds of tables.
Jev ranks every table
One call, every table as an option, a probability for each. The tables are compared against each other, not judged one at a time.
AI2SQL writes the SQL
The LLM gets only the top tables. Smaller prompt, fewer wrong joins.
Join the waitlist
Get early access when Jev table selection ships in AI2SQL.
Questions about Jev table selection
What is Jev table selection in AI2SQL?
A step that runs before AI2SQL writes any SQL. Your question and your schema's table and column names go to Jev in a single call, Jev returns a probability for every table, and the LLM that writes the query sees only the top few. It is coming soon; the waitlist gets it first.
What is Jev?
Jev is a decision model from TypeSafe AI, released on 15 September 2026. It does not write text: it returns a choice with a probability for each option. That makes it fast and cheap enough to rank a whole schema in one call.
How well does it pick the right tables?
On a 56-table test schema with three questions of known answer, run twice each, one Jev call ranked every required table in the top 4. Asking a separate yes/no question per table buried a required table as deep as rank 46, at about ten times the tokens.
Is Jev deterministic?
Close to it. In our tests, five identical calls returned the same answer in 12 out of 12 cases, and the probabilities moved by about 0.02. An item sitting right on a threshold can still flip between runs, which is why AI2SQL keeps the top few tables rather than cutting at a fixed probability.
How accurate is Jev?
It depends on the task. On a different test, triaging 290 GitHub issues against GitHub's own records, Jev was right 97% of the time (119 of 123) when it was at least 80% confident. For table selection, the result above is the one that applies.
Does it work with large schemas?
Yes, that is the point of it. A single Jev question holds up to 255 options, so a schema with more tables than that is narrowed first and then ranked.
What gets sent to Jev?
Your question and your table and column names, so it can rank them. Row data is not sent.
Is AI2SQL affiliated with TypeSafe?
No. AI2SQL uses Jev through TypeSafe's public API.
Last updated . Jev is a model by TypeSafe AI. AI2SQL is not affiliated with TypeSafe. More on Jev: what is Jev · Jev for SQL