pg-jev README

The project's README from github.com/realZachi/pg-jev. As one markdown file: README.md.

pg-jev — ask your Postgres tables questions in plain language

jev — ask your Postgres tables questions in plain language

CI PGXN License Website

Write the condition the way you would say it. Postgres does the rest.

jev lets you filter, rank and classify rows with plain-language conditions. Every row is judged by TypeSafe's Jev, a System One model that returns calibrated probabilities instead of generated text. No index, no embeddings, no vector column.

Website: pgjev.com

CREATE EXTENSION jev CASCADE;

SELECT * FROM people WHERE jev(people, 'the name is European');

SELECT subject, jev_prob(tickets, 'the customer is angry') AS p
FROM tickets ORDER BY p DESC LIMIT 20;

SELECT jev_choice(tickets, 'which team should handle this?',
                  ARRAY['billing', 'technical', 'security', 'sales']) AS team, count(*)
FROM tickets GROUP BY 1;

SELECT name, jev_score(products, 'how luxurious is this product?',
                       ARRAY['budget', 'mid-range', 'premium', 'luxury']) AS luxury
FROM products ORDER BY luxury DESC;

jev() is an ordinary boolean function, so it composes with everything else in SQL: AND age > 40, joins, GROUP BY, LIMIT, ORDER BY jev_prob(...).

How it works

  1. jev(table, 'condition') receives the row as a composite value. The first call for a table + condition starts a read-ahead that streams the table in physical order (TID range scans; OFFSET pages for views). jev.max_prefetch_rows limits how far it searches for a requested row and how many skipped rows it retains; it does not limit the answer cache or total session memory.
  2. Rows are packed jev.batch_size (20) per request into one shared state ({"condition": ..., "rows": [...]}) with one yes/no Noul question per row. Jev evaluates all questions over one state in parallel, which amortises the ~270-token request overhead (about 435 tokens for one row alone vs 175 per row in batches of 20).
  3. Up to 2 × jev.concurrency requests are in flight over persistent HTTPS connections, and every row is answered as soon as its batch returns, so a LIMIT stops the read-ahead after the in-flight window, and rows that cheaper predicates filter out before jev() runs (WHERE age > 60 AND jev(...)) are skipped rather than judged.
  4. Answers are cached per row content and question for the backend session, so re-running, changing the threshold or sorting by probability is free while those answers remain cached. The cache can grow with unique row/question pairs until jev_cache_clear() clears it or the backend session ends. Rows from a subquery or CTE (anonymous record type) can't be read ahead and are judged one request at a time; put jev() on base tables or views when you can.

Measured on a 2,000-row table from Europe (~190 ms to the API): first run ≈ 3.5 s in 100 requests, ≈ 296k input tokens, ≈ $0.012; second run ≈ 50 ms; LIMIT 3 on a new condition ≈ 0.6 s. A new condition in a session that still holds its pooled connections (idle for less than jev.keepalive) takes ≈ 2.3 s: the first request on each fresh connection is the slow one. Version 0.1.0 needed 8.5 s (and 338k tokens) for the full query and 8.4 s for the LIMIT.

Why 20 rows per request

Jev has to find rows[i] by position in the array, and that gets unreliable in long arrays. Against ground truth from structured columns (job title, EU membership, a phrase in a free-text field; 400 rows each), batches of 1–20 rows were 100 % correct, batches of 40 were 92–98 % and batches of 80 were 77–94 %. Wider rows (1,000 characters) made no difference at 20. Naming rows instead of indexing them did not help. Batches of 20 cost 4 % more tokens than batches of 40 and are just as fast, because a request's latency barely depends on its size.

Install

Requirements: PostgreSQL 14–17 with plpython3u (package postgresql-plpython3-NN on Debian/Ubuntu, included in the EDB and Postgres.app builds), a superuser, and a TypeSafe API key from https://console.typesafe.ai. Managed hosts that withhold superuser or plpython3u (Supabase, Neon, RDS, …) cannot run it; see Where it runs.

With an AI agent (easiest)

The repo ships an agent skill on skills.sh. Install it into your project and tell Claude Code, Codex, Cursor or any other skill-aware agent to finish the job:

npx skills add realZachi/pg-jev

Install pgjev on this server and set it up.

The agent runs a preflight (PostgreSQL version, plpython3u, superuser), pgxn install jev or make install against the right pg_config, CREATE EXTENSION jev CASCADE, places the API key and runs a smoke test. Afterwards it also knows how to write cost-conscious jev() queries ("find the tickets where the customer threatens to cancel") and to explain what pgjev can do. The docs are readable as Markdown for agents too: append .md to any page under https://pgjev.com/docs (see For agents).

From PGXN

pip install pgxnclient       # once; also available as `pgxn-client` in Debian/Ubuntu and Homebrew
pgxn install jev             # downloads the release from pgxn.org and runs `make install` against pg_config on PATH
psql -c "CREATE EXTENSION jev CASCADE"

Use pgxn install jev --pg_config=/path/to/pg_config (or sudo pgxn install jev) when the server's pg_config is not on PATH or the extension directory is not writable.

From source (PGXS)

git clone https://github.com/realZachi/pg-jev.git && cd pg-jev
make install            # uses pg_config on PATH; or: make install PG_CONFIG=/path/to/pg_config
psql -c "CREATE EXTENSION jev CASCADE"   # superuser required (plpython3u is untrusted); CASCADE creates plpython3u

Docker

docker build -t pg-jev .                       # add --build-arg PG_MAJOR=17 for another major
docker run -d -p 5432:5432 -e POSTGRES_PASSWORD=pw -e TYPESAFE_API_KEY=your-key pg-jev
psql postgres://postgres:pw@localhost/postgres -c "CREATE EXTENSION jev CASCADE"

API key

Either export TYPESAFE_API_KEY in the environment of the PostgreSQL server process, or set it per session or per role:

SET jev.api_key = 'your-key';
ALTER ROLE analyst SET jev.api_key = 'your-key';   -- persistent, per role

Local Jev-compatible servers

jev.api_url can point at any server that speaks the same POST /v1/systemone contract, so rows never leave your network and there is no per-row billing. Two such servers are stuntd (Laya behind the Jev API; it also records answers and trains a head per question, so a table's decisions can move local over time) and laya-server:

SET jev.api_url = 'http://127.0.0.1:8787/v1/systemone';   -- stuntd's default address

An API key is only required for *.typesafe.ai hosts. For any other host the request is sent without an Authorization header when no key is set, and with Bearer <key> when one is. The token counts and cost estimates in jev_stats() and the notices use TypeSafe's list price, whatever the server.

Functions

Function Returns Purpose
jev(row, condition [, threshold]) boolean WHERE predicate. Threshold: argument → jev.threshold → 0.5
jev_prob(row, condition) float8 Probability 0..1 that the row satisfies the condition
jev_score(row, question, levels text[]) float8 Probability-weighted position on ordered levels (0 .. n-1)
jev_score_norm(row, question, levels) float8 Same, normalised to 0..1
jev_choice(row, question, options text[]) text The most likely option for the row
jev_confidence(row, question, kind, options) float8 Confidence of a score/choice answer
jev_eval(row, question, kind, options) jsonb Full raw answer (probabilities, legend, confidence)
jev_stats() jsonb Requests, tokens, estimated cost, cache hits, in-flight requests and pooled connections for this session
jev_cache_clear() void Forget cached judgments
jev_version() text Extension version

row is the table alias itself (jev(people, ...)) or a subquery alias.

Settings

All settings are plain GUCs: SET jev.<name> = ..., ALTER ROLE ... SET, ALTER DATABASE ... SET, or postgresql.conf.

Setting Default Meaning
jev.api_key env TYPESAFE_API_KEY TypeSafe API key. Optional when jev.api_url is not a *.typesafe.ai host
jev.model jev-latest Model name or pinned version such as jev-1.13.0
jev.threshold 0.5 Probability at which jev() returns true
jev.batch_size 20 Rows per API request. Accuracy drops measurably above ~20–25 (see above)
jev.concurrency 16 Parallel API requests; up to twice that many are queued ahead of the executor
jev.max_prefetch_rows 5000 How far read-ahead searches and how many skipped rows it retains; not a bound on the answer cache or total session memory
jev.notices on Emit a progress NOTICE per finished request and a summary per table with request count, tokens, estimated cost and time
jev.api_url https://api.typesafe.ai/v1/systemone Endpoint (proxies, mocks, local Jev-compatible servers such as stuntd at http://127.0.0.1:8787/v1/systemone; see above)
jev.timeout 30 Seconds per API request. Waits are interruptible: statement_timeout and cancel requests apply within 250 ms
jev.keepalive 600 Seconds a pooled API connection may sit idle before it is reconnected. The first request on a fresh connection costs a TLS handshake plus, measured, up to 1.5 s of server-side setup, so keep connections alive across queries; TCP keepalive probes catch silently dropped ones
jev.max_rows_per_statement 0 (off) Abort a statement that would send more rows than this to the API. Spend guard for shared deployments
jev.max_chars_per_statement 0 (off) Same, for characters of row data

Writing good conditions

Jev answers the question you wrote, literally. A few things that help (more in the TypeSafe docs):

Caveats

Development

make docker-test                 # builds test/Dockerfile and runs the regression suite (PG_MAJOR=16 by default)
make docker-test PG_MAJOR=17

Locally with a running server and pg_config on PATH:

make install
python3 test/mock_api.py &       # deterministic stand-in for the TypeSafe API
make installcheck                # pg_regress, tests in test/sql, expected output in test/expected

The regression tests never call the live API. To try the real thing, SET jev.api_key and run any query.

See CONTRIBUTING.md and docs/PUBLISHING.md for release steps.

License

PostgreSQL License. Jev and TypeSafe are trademarks of their respective owners; this project is not affiliated with TypeSafe.

pg-jev (realZachi/pg-jev)

Searching Database…

1.2K ·GitHub
Pg-jev
Что это?pg-jev (jev) — расширение PostgreSQL, позволяющее фильтровать, ранжировать и классифицировать строки таблиц по условиям на естественном языке вместо SQL-предикатов. Каждую строку оценивает Jev от TypeSafe — модель, возвращающая откалиброванные вероятности, а не сгенерированный текст.
Это отдельное приложение или оно живёт внутри Postgres?Это нативное расширение PostgreSQL, написанное на plpython3u и добавляющее новые SQL-функции вроде jev() и jevprob(), а не отдельное приложение или сервис.
Как установить jev в свою базу данных?Проще всего через AI-агента: установите скилл с skills.sh и попросите Claude Code, Codex или Cursor настроить pgjev на вашем сервере. Иначе установите из PGXN командой 'pgxn install jev --pgconfig=/path/to/pgconfig', соберите из исходников с PGXS или используйте предоставленный Docker-образ.
Что нужно моему серверу Postgres, чтобы это заработало?PostgreSQL 14–17 с расширением plpython3u, доступ суперпользователя и API-ключ TypeSafe (или самохостящийся Jev-совместимый сервер вместо ключа).
Работает ли это на любом управляемом Postgres-хостинге, и какие есть подводные камни?Нет — jev() выполняет полный скан: каждая строка, которую проверяет исполнитель, отправляется в API, поэтому стоимость и задержка растут со сканом (LIMIT, более дешёвые предикаты и кеширование всё это снижают). Управляемые хостинги, не дающие суперпользователя или plpython3u, вроде Supabase, Neon или RDS, не смогут запустить его вовсе.
Покидают ли данные моих строк мою сеть?По умолчанию содержимое строк отправляется в сторонний API для оценки, поэтому не стоит применять его к данным, которые нельзя раскрывать — если только jev.apiurl не указывает на самохостящийся Jev-совместимый сервер: тогда строки никогда не покидают вашу сеть.
В чём разница между jev() и jevprob()?jev() возвращает булев WHERE-предикат, оцениваемый по порогу, а jevprob() — сырую вероятность от 0 до 1 того, что строка удовлетворяет условию, что удобно для изучения распределения перед выбором порога.
Почему он батчит ровно 20 строк на API-запрос?Тестирование по эталонным данным показало, что батчи из 1–20 строк были верны на 100%, падая до 92–98% при 40 строках и 77–94% при 80, потому что Jev должен находить каждую строку по позиции, а на длинных массивах это становится ненадёжным.
Сколько реально стоит выполнение запроса?На таблице из 2000 строк первый прогон стоит около $0.012 за 100 запросов (~296 тыс. входных токенов) и занимает ~3,5 секунды; повторный прогон того же условия с кешем падает примерно до 50 мс, поскольку ответы кешируются по строке и вопросу на время сессии.
Можно держать все свои данные на своей инфраструктуре?Да — укажите jev.apiurl на самохостящийся Jev-совместимый сервер вроде stuntd или laya-server: они говорят по тому же API-контракту и держат строки в вашей сети без построчной тарификации.
What are similar projects?coral, oxynote, ego-lite, vicoa, open-steps, im-not-ai, natural, claudecodeui
Which topics is it in?classification, extension, natural-language-processing, sql, nlp, shell, cursor, skills, postgresql, local-first, database, plugin
Creator
realZachi
Size
0.2 MB
Stars
1.2K
Updated
CommitOct 06Releasev0.2.1Oct 03
Age
1mo
Commits · 90 Days