Are the benchmark numbers real?

They are live jev_stream Choice runs against the real API and jev-1.13.0 on DuckDB 1.5.5 and macOS arm64, reported as medians of three complete queries including execution, network, and inference. Batch 25 reached 474 rows per second on 100 rows and batch 100 reached 1,943 rows per second on 1,000 rows, but the fixed 12-template corpus tests throughput and results vary with input, network, and load.

Why not just use Python UDFs?

This batches and streams inference directly from SQL without Python UDF registration or a separate inference server, linking pinned static libraries so the same binary also loads from Python hosts.

What happens to keys, NULLs, and bad types?

A matching DuckDB secret takes precedence over TYPESAFE_API_KEY and keys never appear in results or cache keys. Any SQL NULL yields NULL without parsing or requesting, while non-finite numbers, top-level scalar numeric or boolean states, and unsupported types without to_json conversion are rejected.

Which DuckDB builds can I load?

Use the build matching your version and platform under build/extension/jev/jev.duckdb_extension, with prebuilt archives published after all eight version and platform builds pass. Local unsigned development requires duckdb minus unsigned or allow_unsigned_extensions in Python.

What functions does the extension add?

jev for a boolean Noul predicate at an explicit threshold, jev_noul, jev_choice, jev_score, jev_eval for mixed questions, and jev_stats for process-level counters. Each returns typed fields plus model and cache_hit.

How do I load the development extension?

Run duckdb -unsigned, then LOAD '/absolute/path/to/build/extension/jev/jev.duckdb_extension'. The Python host loads the same native binary with allow_unsigned_extensions true, and the built artifact is jev.duckdb_extension.

How are credentials supplied?

A scoped DuckDB secret of type jev takes precedence, created with CREATE SECRET (TYPE jev, API_KEY '...', MODEL 'jev-1.13.0'); ENDPOINT is also supported. It falls back to TYPESAFE_API_KEY, and keys never appear in query results or cache keys.

CommunityJev Integrations & Adapters

Duckdb Jev — prasanthj

Native C++ DuckDB extension that runs Noul, Choice, Score, and mixed judgments directly from SQL with batching, concurrent HTTP, and chunked streaming and no Python UDFs.

See highlights

Open github.com

Stars: 9
Language: C++
License: APACHE-2.0
Published by: Community
Author: prasanthj
Last verified: Oct 2026

prasanthj/duckdb-jev

User repository on GitHub

View on GitHub

High-throughput, robust native DuckDB extension for batched and streaming TypeSafe/Jev classification, scoring, and semantic predicates from SQL.

Forks: 2
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

This is one of several Jev listings for the same project. The canonical entry is Duckdb Jev — colliber.

About this project

High-throughput, robust native C++ DuckDB extension for semantic predicates, classification and rubric scoring through TypeSafe/Jev. It batches and streams inference directly from SQL without Python UDF registration or a separate inference server. Real Jev responses on synthetic data; timings are from one local run. Reproduce the animation with vhs examples/terminaldemo.tape. SQL-native judgments: Noul predicates, finite Choice classification, ordered Score rubrics and mixed multi-question evaluation.

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

Tagscommunity jevbest catalog-import c++

Works forJev for Classification

Used inJev for Software & DevTools

Highlights

Packs up to 1,000 judgments per request with bounded concurrency and streams across DuckDB chunks.
Accepts text, JSON, STRUCT, LIST, and ARRAY evidence without exporting through Python or pandas.
Reads a scoped jev secret first, then TYPESAFE_API_KEY, keeping keys out of results and cache keys.
jev_stats reports requests, retries, cache hits, tokens, bytes, and latency with deduplication and LRU reuse.

Watch out

Native binary must match DuckDB version and platform; tested with 1.4.5 and 1.5.6 on macOS and Linux Arm64 and x86-64. Input goes to TypeSafe for remote inference, enable_external_access false blocks requests, and the fixed 12-template corpus measures throughput rather than general accuracy. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

duckdb-jev

DuckDB extension that asks a Jev question about every row in SQL and returns a typed ENUM, DOUBLE, or STRUCT column instead of generated text.