How does IntentSQL use Jev differently from text-to-SQL?

Instead of generating one query, it asks a sequence of bounded Jev decisions about what the request means -- table, column, comparison, operand -- narrowing candidates as established facts flow into later questions, then compiles deterministic SQL from the typed plan and checks whole-request coverage.

What does a refusal look like in IntentSQL?

Unsupported semantics produce an explicit refusal rather than a partial query. For example, top-N-per-group ranking returns 'This request requires unsupported window/ranking semantics; no partial query was executed.' Refusals are counted separately from correct answers in the recorded evidence.

How do I install and run IntentSQL?

Python 3.10 or newer: clone the repo, create a virtualenv, pip install ., then run .venv/bin/intent-sql serve on http://127.0.0.1:7862; a Docker image is also provided, publishing the port on localhost only.

What does one question cost to run?

About two model calls and 8,000 input tokens per question, roughly $0.0004, or $0.042 per million input tokens on TypeSafe Jev. INTENTSQL_BUDGET_USD sets a spending cap, and with INTENTSQL_SPEND_LEDGER it holds across runs.

Does IntentSQL touch my original database?

No. It opens a read-only connection and works on a copy of each database, so the original file is never modified, and requests to change data are refused.

Where are my model keys stored?

Server-side in an owner-only settings file, sent only to their own provider. They are never returned to the browser or written to logs and reports.

How accurate is IntentSQL on real questions?

The README reports 93.3% (42/45) on a held-out clinic schema, 78.0% on 600 Spider test questions, and 87.5% when it is confident, meaning roughly one in eight confident answers is still wrong.

CommunityJev Integrations & Adapters

IntentSQL

Alpha playground that answers plain-English questions over SQLite by composing small Jev decisions -- table, column, comparison and operand selection -- into inspectable typed plans, deterministic SQL and previewed guarded writes. Bundles Cyberchase, DESE and Moneyball databases.

See highlights

Open github.com

Stars: 3
Language: Python
License: MIT
Published by: Community
Last verified: Oct 2026

Amine-LG/IntentSQL

User repository on GitHub

View on GitHub

Natural-language SQLite interface and experiment with Jev / System One decision models: inspectable query plans, deterministic SQL, and guarded writes.

Forks: 1
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

The query, made clear. Ask a SQLite database a question in plain English. IntentSQL shows what it understood, every decision behind the SQL, the read-only query it ran, and the result. A real answer from TypeSafe Jev (jev-1.13.0): two model calls, 10,678 input tokens, about $0.0005. Jev's check was unsure this time, so the answer is flagged Uncertain and the other reading is offered beside it.

Quoted from the project's README. Read the full README.

Tagscommunity sqlite text-to-sql natural-language

Works forJev for Extraction Jev for Classification Jev for Decision gating

Used inJev for Data & Analytics

Languages: Python
Platforms: Browser SQLite
Capabilities: Open sourceRuns locallyRequires API key

Highlights

Recorded evidence: 80/80 on the internal Alpha v2 development corpus and 30/40 on a frozen eligible Spider subset, with refusals counted separately.
Reads support one declared foreign-key hop, aggregates, grouping and HAVING; top-N-per-group ranking is explicitly refused rather than approximated.
Writes compile, preview on an isolated copy, pass a semantic vet, then require explicit confirmation; a 100-change cap includes cascades and trigger effects.
Runs locally on Python 3.10+ at 127.0.0.1:7862 or in Docker with a named volume; bundled databases come from Harvard CS50's SQL course.
The Jev preset uses https://api.typesafe.ai/v1/systemone with model jev-latest; OpenJEV and other compatible endpoints are configurable.

Watch out

v0.1.0-alpha.1 is a local, single-user experiment: accepted interpretations are not proof of correctness, and the recorded evaluation snapshots are not general accuracy estimates. Code is MIT, while the bundled CS50 databases are CC BY-NC-SA 4.0.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

Jev RAG

Local-first knowledge search with seven selectable pipelines: BM25 plus Jev reranking is the default vector-free path, plus agentic, embedding-hybrid, taxonomy, gate and line-search modes. Reranks up to 30 candidates in one batched set of Jev Noul relevance questions and streams cited answers from a configurable model.

Jev Integrations & Adapters#community#rag#retrieval

Community

aifabrice/jev-ragPython

sqlite-jev

Loadable SQLite C extension and Python package that exposes Jev Noul, Choice, and Score judgments as SQL scalar functions and a jev_rows virtual table which batches up to 40 rows into one shared state and joins typed results back to the source by rowid.

Jev Integrations & Adapters#community#sqlite#sql

Community

mgaitan/sqlite-jevC

SignalChain

Data-cleaning framework that asks Jev Choice questions to classify each CSV column into a 13-code type vocabulary and each dataset into one of six scenes, then executes every write locally; an optional System 2 LLM handles the default path or escalations.

Jev Integrations & Adapters#community#data-cleaning#pandas

Community

Cogniform/AI-decision-makerPython

GPTCache

8.2k

A semantic cache for LLM queries, fully integrated with LangChain and llama_index. Zilliz's cache integrates TypeSafe Jev Noul checks to evaluate cache hit freshness and time-dependent query validity, cutting repeated requests and API cost.

Jev Integrations & Adapters#community#semantic-cache#langchain

Community

zilliztech/GPTCachePython

jev4pg

224

A Jev harness for PostgreSQL: 41 semantic operators plus a native Rust extension evaluate row-level predicates with Noul questions and keep the returned probabilities for explicit decision thresholds, leaving joins and aggregation to SQL.

Jev Integrations & Adapters#community#postgresql#nl2sql

Community

Sheltercosmo/jev4pgPython

Laser

Focus mode for Omarchy and Hyprland that judges each screen against the task you declared and escalates from a red reticle to a window haze and, as a last resort, closing the drifting browser tab. Each check is one Jev on-task classification, with screenshots OCR'd locally and never uploaded.

Jev Integrations & Adapters#community#focus#linux

Community

iYassr/omarchy-laserPython

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai