jevsql
Text-to-SQL experiment where Jev answers typed questions about the request and code assembles the query, so the model never writes SQL.
Overview
This jevsql is a weekend experiment in text-to-SQL where the model never writes SQL. Instead of asking a generative model to produce a query and hoping it parses, the system asks Jev typed Choice, Score and Noul questions about the user's natural-language request and the schema, such as which table, which columns, which filters and which aggregation apply, and code assembles a valid SQL query from the answers.
The README spends much of its length explaining how Jev differs from a generative LLM: typed values with no parsing errors, calibrated confidence per answer, many questions at about the latency of one, and much lower price, while being unable to compose anything itself. It argues for a cascade in which Jev decides on every request and a generative model is used only for the fraction that needs one. The details of the SQL assembly and any results are beyond the README excerpt. It is aimed at developers curious about structure-safe text-to-SQL.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which table does the request refer to?(inferred) | — |
| Choice | Which column, filter or aggregation does the request need?(inferred) | — |
| Yes / no | Does the request need this condition?(inferred) | — |
How it works
- User asks a data question in natural language
- Code builds typed questions from the schema
- Jev answers in one parallel pass
- Code assembles a SQL query from the answers and runs it
Strengths
- Query structure valid by construction
- Calibrated confidence per component
- Cheap and fast compared with generative text-to-SQL
Caveats
- Weekend experiment
- Limited to query shapes the code can assemble
- Results and coverage not established in the README excerpt
At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- script
- Has code
- yes
Use cases and integrations
How curated lists describe it
Text-to-SQL where the model never writes SQL — typed, calibrated decisions (TypeSafe Jev) + code-assembled queries
logicrw/awesome-jev-projects
Frequently asked questions
Does Jev write the SQL?
No. Jev answers typed questions and code assembles the query.
Why do it this way?
Assembled queries cannot have syntax errors and each component has a calibrated confidence.
Is it production-ready?
No, the author describes it as a weekend experiment.
Related projects
PostgreSQL toolkit with 41 Jev-powered semantic operators and natural-language-to-SQL, leaving joins and arithmetic to SQL.
ProjectbetaPythonResume screening where an LLM writes the questions once per job and Jev answers them for every resume, quoting the evidence line.
ProjectproductionTypeScriptClassifies tax PDF pages into 261 IRS forms and 7 page kinds with one Jev request per page and a confidence gate.
ProjectproductionTypeScriptRust tool that asks Jev which PDF pages genuinely need OCR, extracting text pages locally and sending only the rest to OCR.
ProjectbetaRustFills Japanese government PDF forms by asking Jev which profile item each unlabeled field is; personal values stay in the browser.
ServicebetaJavaScriptDocker service that classifies the Paperless-ngx inbox with Jev (or local Clef via Ollama), setting confident fields and queueing the rest.
Projectbeta