IntentSQL
Natural-language SQLite interface where code proposes options, Jev chooses among them, and a compiler builds deterministic SQL.
Overview
IntentSQL lets you ask a SQLite database a question in plain English without ever letting a model write SQL. Code reads the database and question, grounds the values you mention to tables and columns, and builds every reasonable option for which column to show, which value filters which column, and how to group, sort or join. A decision model answers these small multiple-choice questions in one parallel round, the answers are assembled into a typed plan described back in plain English, and a compiler turns the plan into parameterized SQLite.
Readings that differ by one decision and give different results are run too, and the model checks each plain-English reading against the question; uncertain answers are flagged with alternatives, using confidence calibrated on measured accuracy. Queries run read-only with row and time limits, and write requests are refused. A decision timeline shows each exact question, the chosen option and its probabilities. The README shows a real Jev answer in two calls at about $0.0005, and supports Jev, OpenJEV, a local LAYA adapter or a custom System One endpoint.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which column should be shown for this question?(inferred) | — |
| Choice | Which value filters which column?(inferred) | — |
| Yes / no | Does this plain-English reading match the user's question?(inferred) | yes · no |
How it works
- Ground mentioned values to tables and columns in code
- Build all candidate options for each plan decision
- Jev answers the bounded questions in one parallel round
- Compile the typed plan to parameterized read-only SQL
- Run alternative readings, verify, and flag uncertain answers
Strengths
- Model never writes SQL, so queries are deterministic and parameterized
- Every decision and probability is visible
- Calibrated confidence flags uncertain answers with alternatives
- Read-only execution with limits; writes refused
Caveats
- SQLite only; supported features are listed with exclusions
- Released as an experiment alongside a decision lab
At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- web app
- Has code
- yes
Use cases and integrations
How curated lists describe it
Natural-language SQL: turns a question about a SQLite database into a sequence of small Jev decisions instead of one generated query, released as an experiment alongside its decision lab.
yibie/awesome-jevNatural-language SQLite interface and experiment with Jev / System One decision models: inspectable query plans, deterministic SQL, and guarded writes.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide in IntentSQL?
Small multiple-choice questions about which column, filter, grouping or join you meant, plus a check of each reading.
Can it modify my database?
No. Queries run on a read-only connection and change requests are refused.
Which backends are supported?
Jev, OpenJEV, a local LAYA adapter, or a custom System One endpoint.
Related projects
SQLite and Node library adding Jev judgments to SQL for matching, evidence picking, ranking, review queues and semantic CI checks.
ProjectbetaJavaScriptMeasures whether SQL ORDER BY over Jev probabilities gives a defensible ranking: inversions, ordinality, calibration, wording and ties.
BenchmarkresearchPythonGitHub Action that triages issues with Jev, applying label, spam, needs-info and duplicate tags only above set thresholds.
ProjectbetaTypeScriptOne YAML task to classify, route, score or verify text with rules, Jev, NanoJev or an OpenAI-compatible model, with cost records.
CLIbetaPythonMultiplayer GitHub triage dashboard where Jev answers kind, severity, urgency, duplicate and next step per issue.
AppbetaCLI, MCP server and skill for issue triage where Jev predicts issue kind, affected module and whether information is sufficient.
MCP serverprototypePython