SOLO

Reorders rows and fields so a prefix-caching backend reuses more work when a decision model judges every record in a large table; ships a vLLM backend for the open AutoTrust JEV-9B model and works with your own.

View on GitHub →

Checked, not fully reviewed

Reorders rows and fields so a vLLM-served open decision model reuses cached work. Checked, not fully reviewed: correct on CPU; needs a GPU; never calls Jev.

Good for

Running one yes/no or pick-one question over every row of a large table on your own GPU, where rows share long policy or account text
Seeing why a layout should help before you rent a GPU: `explain()` reports the shared prefixes in bytes and says what they do not predict

Watch out for

It does not call TypeSafe's Jev. Both backends are open models served by vLLM on an NVIDIA GPU (the reference setup is a 24 GB RTX 4090)
The speedups depend on prefix caching on your own server; the README says they disappear with it off

Tested Oct 18, 2026 at c5bed0511f3c · Python 3.12 in Docker on 2 CPU cores, no GPU and no model server; its test suite and its offline layout example

Checked, not fully reviewed. We installed SOLO from source on a CPU, ran its test suite and its offline example, and read both backends. We have no NVIDIA GPU, so we did not run a model, and the README’s throughput and accuracy figures (measured on an RTX 4090) are the author’s own. No Jev calls.

SOLO asks one natural-language question of every record in a table, a DataFrame or a JSONL file. Before sending the records, it reorders them. Fields that many rows share, such as a refund policy or an account tier, move to the front. Rows with the same values sit next to each other. A server with prefix caching can then reuse the work it did on the shared part. Answers come back in the original row order.

The layout part works without a model. The test suite passed: 76 tests, with 8 skipped because they need the optional plotting and tokenizer packages. In the offline example, four support tickets came back with policy and account ahead of id and grouped by policy. Every field was still in every record, which is the README’s promise: the model sees the complete record, not a deduplicated one. Answers mapped back to the input order. explain() reported the share of adjacent bytes that repeat as rising from 4% to 34%. It also labels that number as describing structure, not tokens, cache hits or speed, which is an honest caveat.

The model part needs a GPU, and is not Jev. One backend sends a one-token prompt to vLLM’s /v1/completions for AutoTrust JEV-9B, an Apache-2.0 decision model on Qwen3.5-9B. The docs ask for a 24 GB GPU and about 35 GB of disk. The other talks to vllm-jev serving Open-Jev-2B. Neither speaks TypeSafe’s /v1/systemone, and the project gives no way to point it at the hosted Jev. It is a tool for running open decision models in Jev’s style on your own hardware.

MIT. 804 stars, most of them gained this week.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Oct 18, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.

More in Data & Observability

pg-jev

★ 1.2k▲ 24

realZachi/pg-jev

PostgreSQL extension to filter, rank, and classify rows with plain-language conditions.

Shell

DocJev

★ 546▲ 7

jerryjliu/docjev

Classifies and splits PDF, DOCX and PPTX with Jev and local OCR, asking one typed question per page and per boundary in a single request. Ships the manifest, per-call records and error analysis behind its benchmark.

PythonReviewed

tax-doc-classifier

★ 515▲ 1

kyotofin/tax-doc-classifier

Classifies tax-document pages into IRS forms and page kinds with one Jev request per page, driven by a JSON file of form descriptions.

TypeScriptReviewed

Get new Jev projects every week

New Jev releases, pricing changes, and the best new projects, once a week. No spam; unsubscribe anytime.