jev4pg
Natural language to SQL for PostgreSQL, plus 41 Jev operators. On the project's BIRD table, Jev alone scores below the LLM.
Overview
jev4pg (github.com/Sheltercosmo/jev4pg, jev4pg.com) is an Apache 2.0 PostgreSQL project from Shelter Zhou (Sheltercosmo). You ask a question in English or Simplified Chinese and get SQL you can read, plus a workspace, an HTTP API, and durable query jobs in release v0.7.0. Older names include JevSDSQL and jevsd-pg.
The repo ships 41 Jev operators for filtering, extraction, ranking, matching, and verification. Explainable embeddings store named questions and answer probabilities. Evidence caching keeps raw observations when a threshold changes. VALUE, UNKNOWN, and NOT_EVALUATED stay separate from execution failures.
On the README BIRD Challenging snapshot (100 questions, 11 databases), Jev 1.13.0 alone matched 20 of 99 scorable answers (20.2%). GPT-5.6 Terra matched 39 of 99 (39.4%). Jev plus GPT-5.6 Terra matched 34 of 99 (34.3%). Jev alone was cheaper; the LLM baseline matched more answers. Methodology limits are in the FAQ and sourced metrics below.
Built for data and backend teams that already trust PostgreSQL and want semantic steps with a review path before anything is written.
Problem: People want to ask a Postgres database a question in ordinary language and still see the SQL, the joins, and the sums, with a clear mark when a judgment was skipped.
Built for: Data and backend teams who already run PostgreSQL and want English or Simplified Chinese questions, a large set of Jev operators, and a review step before anything is written.
First indexed on Jev Directory: 2026-10-06
Creator and team
- Name
- Shelter Zhou
- Handle
- @Sheltercosmo
“Natural language to SQL and semantic operators for PostgreSQL.”
How Jev is used
- Role in the product flow
- Plan SQL from a natural-language question, and run Jev operators for the semantic steps Postgres does not compute
- Primitives
- ChoiceScoreNoul
- State in
- A question plus the tables you attach. Operator calls take text or dataset rows and a typed question. The native scan preview takes a source SELECT and a JSON question map, such as a noul on a column.
- Decision out
- A SQL proposal you can inspect, or an operator result that keeps VALUE, UNKNOWN, and NOT_EVALUATED apart from execution failures. Exact math stays in PostgreSQL.
- Clone the v0.7.0 tag, run python deploy/configure.py, then docker compose up. The README asks for Python 3.11+ and Docker Compose v2.
- Open the local English workspace at /ask/en or the Simplified Chinese workspace at /ask/zh.
- POST /ask with planner_mode set to jev (no LLM generation) or hybrid (Jev picks context, then an LLM proposes SQL).
- Use the 41 operators for filter, extract, rank, match, and verify work. The operator catalog is sdd/operators/manifest.json.
- Optional native preview: Rust jev_native functions on PostgreSQL 17 for Linux, including parallel plans and probability embeddings.
jev4pg is the planning and operator layer. PostgreSQL still does joins, windows, and exact arithmetic. A jev planner mode builds SQL with no LLM generation call. Hybrid mode lets Jev pick schema and evidence, then an LLM writes a proposal that Jev reviews. You can correct the saved interpretation and reuse it. pg-jev (https://pgjev.com?ref=jev.aitools.fyi) is a separate Postgres extension with one jev() predicate per row in a WHERE clause. The native Rust path is a preview in v0.7.0 for PostgreSQL 17 on Linux. The project guide is https://jev4pg.com/guide/. Cite the jev4pg README and docs/benchmarks/BIRD_CHALLENGING_100.md for the scores.
Sourced performance claims
- README table, BIRD Challenging, 100 questions and 11 databases: Jev 1.13.0 matched 20/99 (20.2%), median 8.55 s, estimated $0.389 per 100 attempts. GPT-5.6 Terra matched 39/99 (39.4%), median 8.68 s, $3.262. Jev plus GPT-5.6 Terra matched 34/99 (34.3%), median 14.23 s, $2.839. Jev alone scored lower than the LLM.Source: github.com/Sheltercosmo/jev4pg README, BIRD Challenging section
- Jev planned with 0 LLM generation calls at 88.1% lower estimated token cost than the LLM baseline. Hybrid used 20.8% fewer LLM input tokens. The LLM baseline matched more answers, and hybrid took longer. 1 unavailable reference leaves 99 scorable questions. Medians cover 94 cases with up to 3 cases in flight. Costs use frozen accounting rates. 1 hybrid request has unreported usage. Measured on the frozen Python planner on 23 September 2026. Not an official leaderboard score, and not a measurement of v0.7.0 or the Rust preview.Source: github.com/Sheltercosmo/jev4pg README, BIRD Challenging section
- Methodology note: hybrid matched 6 answers the LLM missed and missed 11 the LLM matched. Both matched 28. The run does not establish that hybrid is more accurate or faster. p95 times were 26.80 s, 38.85 s, and 54.16 s. Complete application results matched 20, 38, and 34 references. 1 otherwise correct LLM query passed the application's 1,000-row return cap.Source: github.com/Sheltercosmo/jev4pg docs/benchmarks/BIRD_CHALLENGING_100.md
- The README and sdd/operators/manifest.json describe 41 Jev operators. GitHub release v0.7.0 was published on 4 October 2026.Source: jev4pg README, sdd/operators/manifest.json, GitHub releases
- Public GitHub repo Sheltercosmo/jev4pg had 143 stars and 2 forks on 2026-10-06. License Apache-2.0. Created 2026-09-23. Language Python. Homepage https://jev4pg.com.Source: GitHub API 2026-10-06
Features and stack
Features
- Natural-language questions in English and Simplified Chinese, with the SQL left visible
- 41 Jev operators for filtering, extraction, ranking, matching, verification, and workflow steps
- Workspace, HTTP API, and asynchronous jev SQL jobs in release v0.7.0
- Explainable embeddings from named questions and answer probabilities (native preview)
- Evidence caching, controlled retries, and reviewable fallbacks that keep skipped work explicit
- Read-only attachments so authorized tables stay in place
- BIRD Challenging comparison where Jev alone is cheaper and less accurate than the LLM baseline
Stack
- Python 3.11+
- PostgreSQL
- Docker Compose
- FastAPI
- Optional Rust native preview
- TypeSafe or a compatible HTTP endpoint
- Apache-2.0
Pricing: Apache 2.0 code. Inference uses your TypeSafe key or another provider you configure. Hybrid planning also needs an LLM provider. The BIRD dollar figures are frozen accounting estimates in the README, not an invoice.
Links
FAQ
- How is jev4pg different from pg-jev?
- pg-jev is a Postgres extension with one jev() function. You put it in a WHERE clause and it judges one row at a time. jev4pg plans a SQL query from a natural-language question, ships 41 Jev operators, and publishes a BIRD Challenging comparison. The titles on this page stay on jev4pg so the two listings do not compete.
- Did Jev beat the LLM on the BIRD test?
- No. The README table says Jev 1.13.0 matched 20 of 99 (20.2%). GPT-5.6 Terra matched 39 of 99 (39.4%). Hybrid matched 34 of 99 (34.3%). Jev alone was the cheap run: 0 LLM generation calls, and an estimated token cost 88.1% lower than the LLM baseline. Hybrid used 20.8% fewer LLM input tokens and had a slower median, 14.23 s against 8.68 s.
- Do these BIRD numbers measure release v0.7.0?
- The README says no. The table is a local historical comparison of the frozen Python planner on 23 September 2026. It is not an official leaderboard score, and it does not measure v0.7.0 or the Rust preview. 1 missing reference leaves 99 scorable questions. Medians cover 94 cases. The methodology note adds that hybrid was not shown to be more accurate or faster: it matched 6 answers the LLM missed and missed 11 the LLM matched.
- Which release should I install?
- GitHub's latest release is v0.7.0, published 4 October 2026, titled PostgreSQL workspace and durable queries. The README install block clones that tag.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.