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), so memory stays constant whatever the table size.
  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 for the session, so re-running, changing the threshold or sorting by probability is free. 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

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
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 past a cache miss the read-ahead scans to find the requested row, and how many skipped rows it keeps for later requests (memory bound)
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)
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):

  • State the exact condition: 'the customer threatens to leave, dispute a charge, or take legal action' beats 'churn risk'.
  • Keep arithmetic, dates and exact matches in SQL; let the model judge meaning.
  • Look at the distribution with jev_prob() before picking a threshold. Ambiguous cases really do land near 0.5.
  • Send only the columns the judgment needs: create a view with the relevant columns (and any pre-filter) and call jev(view_alias, ...) on the view. Views are read ahead and batched like tables.

Caveats

  • This is a full scan by design: every row the executor asks about goes to the API. Cheaper predicates in the same WHERE run first and their rejects are skipped; a LIMIT stops early; jev.max_rows_per_statement caps spend.
  • Row contents are sent to a third-party API. Do not use it on data you may not share.
  • The cache lives in the backend session (PL/Python GD). Connection pools with many sessions each warm their own cache.
  • plpython3u is an untrusted language: only superusers can create the extension, and functions run with the server’s OS privileges.

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.