Jev4pg

No. Jev judges row-level predicates through Noul questions per row or distinct value, and PostgreSQL handles joins, window functions and aggregation on your existing data.

The Python app provides the workspace, file imports and natural-language planning; the Rust jev_native extension runs semantic functions inside PostgreSQL for direct SQL use without a Python service.

Filtering, extraction, ranking, matching and verification, plus explainable embeddings that represent data through named questions and answer probabilities so similarity stays inspectable. The native preview adds parallel execution and input deduplication, where repeated values with the same context and questions share judgments while every source row is preserved.

Clone the v0.7.0 branch, run python deploy/configure.py to enter your TypeSafe key, then docker compose build and docker compose up -d --wait. It needs Python 3.11+ and Docker with Compose v2. Ordinary SQL requires no model key, Hybrid mode additionally needs LLM configuration, and the workspace opens in English or Simplified Chinese at http://127.0.0.1:8000/ask/.

Write SELECT product, COUNT(*) FROM tickets WHERE SEMANTIC(body, 'The ticket describes an unresolved billing problem.') GROUP BY product ORDER BY COUNT(*) DESC. JEV evaluates the text through Noul questions and PostgreSQL performs the grouping and counting. The README tells you to check decision states and result completeness first, since uncertainty and unevaluated rows stay explicit.

The project is Apache-2.0 with about 214 stars, and release v0.7.0 ships a native extension 0.2.0 preview. Native maintained-feature refresh and semantic write review are still on the roadmap, so the Python engine is the right choice for those workflows. Either version can incur provider usage, and the self-developing semantic layer keeps reviewed definitions and corrections for reuse.

Evidence caching retains raw observations for compatible queries and threshold changes, with source and evaluator identity checks, so moving a cut-off does not re-pay for every judgment. Controlled retries and reviewable fallbacks keep errors, uncertainty and skipped work explicit and preserve legal SQL proposals for correction, while reviewed definitions and corrections build a self-developing semantic layer your team can reuse across reports and applications.