How does it avoid hallucinated extractions?

Jev never writes output text. Code enumerates candidate spans from the document and Jev only chooses which schema class each span belongs to, or none, so offsets always point into the source.

Can I run it without Jev?

Yes, the decider option accepts a chat LLM through OpenRouter for ablation (llm: ); the README says this barely changes quality but costs about 200x more.

What do candidate spans come from?

Deterministic patterns: money, percentages, dates, durations, frequencies, identifiers like CPF/CNPJ, e-mail and URLs, measures, quantities, quotes, proper-noun runs, content n-grams, and any regex your schema declares. Portuguese and English are built in. Segmentation keeps exact sentence offsets, and everything is one typed request per sentence with sentences in parallel.

How does it compare to LangExtract on quality?

It trails, and the README says so. On jx-bench the gap to LangExtract is inside the confidence intervals, but on LeNER-Br and Few-NERD it is clear, with recall high and precision lower. Relaxed span F1 on jx-bench is 84.2 against 88.5 for LangExtract and 92.0 for an LLM JSON baseline, and the optional mention gate raises jevextract to 88.0.

Where does it win outright?

Cost and speed. The benchmark puts it at $1.05 to $1.50 per 1,000 documents with a 0.6 to 1.1 second median per document, against $13 to $28 and 7 to 14 seconds for LangExtract on its default model and $4 and 1.7 to 2.3 seconds on its cheapest. Grounding is also structural rather than checked: doc.text[e.start:e.end] == e.text holds by construction with no alignment step.

Can I rely on identical results across runs?

No. The README reports that Jev's decisions are not deterministic: two full runs agree at a Jaccard of 0.86 to 0.97, and confidences are over-confident with ECE from 0.17 to 0.44, so thresholds need care. It is MIT licensed, Python 3.10+, depends on the official typesafe-sdk and PyYAML, and shipped with one star at verification alongside a bilingual benchmark and a paper.

How is a schema written, and how deep does it go?

A YAML document with a name, an optional min_confidence of 0.6 and a list of classes, each with an id, a description and optional attributes or patterns. Levels can be span-level like party or amount, or sentence-level like obligation, and closed-set attributes cover a party's role or a drug's status, all from one schema with no few-shot examples. Attribute values come from one more call after the spans are accepted.

CommunityJev Tools & CLIs

Jevextract

Grounded information extraction library: deterministic code proposes candidate spans with exact offsets and Jev answers one Choice per span plus sentence-level Nouls, so every extraction is a substring of the source. Ships a bilingual benchmark and paper.

See highlights

Open github.com

Stars: 1
Language: Python
License: MIT
Published by: Community
Last verified: Oct 2026

gabazureus/jevextract

User repository on GitHub

View on GitHub

Grounded information extraction that cannot hallucinate: code proposes spans, Jev decides. An open-source alternative to LangExtract, with a bilingual benchmark and paper.

Forks: 1
Last pushed: Updated Sep 2026

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

About this project

Grounded information extraction that cannot hallucinate. Code proposes the spans, Jev decides. An open-source alternative to LangExtract.

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

Tagscommunity information-extraction nlp benchmark

Works forJev for Extraction Jev for Document processing Jev for Verification

Used inJev for Legal & Compliance Jev for Media & Entertainment Jev for Healthcare & Life Sciences

Languages: Python
Platforms: CLI
Capabilities: Open sourceRequires API key

Highlights

doc.text[start:end] == e.text holds for every extraction by construction; there is no alignment step and no LLM-written output
Benchmark: $1.05-1.50 per 1,000 documents and 0.6-1.1 s median per document versus $13-28 and 7-14 s for LangExtract on Gemini 3.5 Flash
jx-bench relaxed span F1 is 84.2 versus 88.5 for LangExtract and 92.0 for an LLM JSON baseline; the optional mention gate raises it to 88.0
One request per sentence with sentences in parallel; schemas are YAML with class descriptions, closed-set attributes and optional regexes
README reports Jev decisions are not deterministic: two runs agree at Jaccard 0.86-0.97 and confidences are over-confident (ECE 0.17-0.44)

Watch out

MIT license. The README openly reports trailing LangExtract quality on LeNER-Br and Few-NERD, and the project had 1 star at verification; Jev confidence is over-confident, so thresholds need care.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

DocJev

523

Classifies and splits PDFs, DOCX, and PPTX with local page text from LiteParse and category decisions from Jev.

Jev Tools & CLIs#community#python#document

Community

jerryjliu/docjevPython

Translation QA Gate

A localization pre-review gate that pairs source and target strings and judges meaning, terminology, tone, and placeholders as separate questions, so one failing dimension blocks a release without blending scores.

Jev Tools & CLIs#community#python#localization

Community

Skill Scanner

2.6k

Cisco's security scanner for agent skills added a System One analyzer: it sends the full skill text to POST /v1/systemone as one advisory Noul question and records the probability, while being deliberately unable to emit a finding or change a severity.

Jev Tools & CLIs#community#security#agent-skills

Community

cisco-ai-defense/skill-scannerPython

grev

grep, sort, cut and uniq that match by meaning: each line whose Jev Noul clears the threshold survives grev, sibling filters route by Choice and rank by Score, and the output is always your own input, never generated text.

Jev Tools & CLIs#community#cli#coreutils

Community

aurorainfra/grevGo

Edward

Supervisor for unattended coding agents: it tracks cross-turn state and asks one batched Jev Choice whether a trajectory should continue, pause or escalate, while deterministic code owns dangerous-command blocking, budget caps and Ed25519-signed receipts.

Jev Tools & CLIs#community#guardrails#safety

Community

VeridicalTech/EdwardPython

JevSpan

Zero-shot named entity recognition library on Jev: text is split at punctuation, candidate windows are offered as one Choice question per entity type, then verified and boundary-polished by further Choice calls. Types are defined in plain words; no training data or GPU is needed.

Jev Repos & SDKs#community#ner#extraction

Community

lzq-0529/jev-spanPython

From the Jev guides

Original write-ups on what this entry is for.

Jev on the Gateways: Vercel, Workers AI, and OpenRouter

Oct 12, 2026

The request shape barely changes across surfaces; the surrounding contract does. When to call the API directly and when a gateway, a Worker binding, or a router entry earns its extra hop.

Jev Tools & CLIs#gateway#vercel#cloudflare

Read article

Exporting Decision Models: GGUF, ONNX, and Device Runtimes

Oct 10, 2026

Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.

Jev Tools & CLIs#open-models#local#onnx

Read article

Records at Rest: Documents, Inboxes, and Risk Queues

Sep 28, 2026

Jev labels a record that already exists. Code moves, flags, or holds it. How the document, mail, and transfer tools keep the action out of the model.

Jev Tools & CLIs#classification#email#security

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai