First test with Jev (TypeSafe System One): 200 rows, 32 template families, traps — 'yes' rows with "we'll pass", 'no' rows with 'absolutely, happy to meet'. The single Jev question scored 100%. My 12 dimensions scored 98%. I had designed a task my own question could ace.
PatternsSemantic find
Line-by-line search
Rank every line of a document in one request with a Choice, while a Noul checks whether the document contains an answer at all.
- Category
- Patterns
- Also known as
- Semantic find
- Related terms
- 4
- Directory entries
- 33
- Docs
- docs.typesafe.ai
- Added
- 2026-09-24
Definition
The cookbook tags and ranks all 218 lines of GitHub's Terms of Service in a single request. An existence Noul rides along so the system can say no: on an absent answer, one line still scored 0.86 while the exists question read 0.14, which is why both thresholds need tuning.
Choice caps at 255 options, so longer documents use a two-pass window search. The pattern is a model-based alternative to embedding search when the corpus is small enough to fit one request.
Tagspatternssearch
Related terms
Definitions that connect to this one.
Re-rankingScore each query-candidate pair with one question and reorder a shortlist; the CLERC example lifts top-1 accuracy from 5% to 18%.PatternsPrimitive limitsThe hard caps to design around: 255 Choice options, 2 to 10 Score levels, and a 64k context covering state plus questions.PrimitivesNoulThe yes/no primitive: one probability from 0 to 1, where the value is the answer and no separate confidence is returned.PrimitivesRAG passage classificationScore each retrieved passage with four Noul questions, then decide in code which passages reach the answering model.Patterns
From the directory
Semantic search over GitHub's Terms of Service: one request ranks all 218 lines with a Choice while a Noul checks whether the document contains an answer at all, including when it should say no.
Cookbooks & DemosDocs#official#cookbook#search
Official
An independent evaluation of Jev in three retrieval stacks, with recalculated recall, latency, and cost against DeepSeek, GPT-4o-mini, and GPT-5-mini baselines.
Practices & PatternsArticle#community#search#reranking
Community
Describe the startup you are looking for and Jev ranks the pile, asking once per startup whether it is a fit.
Tools & IntegrationsPlayground#community#playground#search
Community
2GitHub stars
An MCP server for Antigravity IDE, Cursor, and Claude Code that uses Jev (jev-latest) to locate code semantically and audit git diffs, replacing broad grep sweeps of entire files.
Tools & Integrations#community#python#mcp
7GitHub stars
Neovim plugin that splits the buffer into functions with Treesitter, asks Jev one plain-language question per function, and fills quickfix with the hits ranked by probability. It judges the unsaved buffer rather than the file on disk.
Tools & Integrations#community#neovim#lua
Community
valentynkit/jev.nvimLua
valentynkit/jev.nvimLua8GitHub stars
Reranking benchmark that gave Jev, Cohere Rerank 4 Pro, ZeroEntropy zerank-2, and DeepSeek the same thirty BM25 candidates across eight English datasets, publishing saved responses, scoring code, and paired-bootstrap intervals. Jev's rubric scored 0.692 nDCG@10 against Cohere Pro's 0.691.
Practices & Patterns#community#python#search
Community
anessbelbati/jev-rerank-benchPython
anessbelbati/jev-rerank-benchPython27 more matching entries in the full directory.
From the community
Posts from builders shipping with Jev right now.
200 rows, 32 template families, and traps, judged in one call
jev(): Postgres WHERE clauses in plain language
I think I just cooked something 🔥 jev(): a PostgreSQL extension that searches your whole database in natural language. No index, no embeddings, just one function. WHERE jev(people, 'could work from home') or WHERE jev(people, 'name sounds european') 129 rows judged in ~1s Show more
@CompleteSkeptic
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x
A real-time ad blocker that classifies every DOM element
I build an undetectable realtime adblocker extension with typesafe It checks every dom element and classifies as ad/non-ad and removes it if true Extremely fun to work with, expecting an incredible shift in how AI is being used in the future
@CompleteSkeptic
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x
Jev plus Astra beats the Ender Dragon for under a dollar
Harness: I made Jev control the player, has access to WASD, space, click, and mouse movements. Astra is the planner that sends instructions to Jev async. Learning: Each time the agent would fail, Astra would add skills as mjs files, that it can draw upon in different scenarios, Show more
Tool calling as classify plus action, back to twelve-factor agents
jev is the best excuse you could possibly have to go re-read 12 factor agents. Tool calling itself can be decomposed into classify+action, if you learn to design ai programs as pipelines that switch breathlessly between classification, structuring data, deterministic code, AND Show more
@dillon_mulroy
i think jev is resonating with devs so well b/c it unlocks so many opportunities for composing ai into systems and products rather than ai _becoming_ the product/system really does feel like it was a missing primitive
Why Jev encodes a state and outputs distributions instead
Jev is one of the more interesting model launches I have seen recently because it asks a very simple question: Why are we using autoregressive LLMs as insanely expensive if statements inside software? The easiest way to think about it is: LLM: text -> generate tokens Show more
@CompleteSkeptic
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x