When will visitors notice a difference from plain keyword search?

When visitors phrase questions instead of keywords. The keyword pass ties classic libraries on exact matches, while Jev lifts intent queries that keyword overlap misses.

Does this lock me into another search dependency?

Yes. One shadcn add lands the component, hook, scorer, server handler, and route as your files. No wrapper package stays in your dependency tree, though you still need a TypeSafe key.

How does one Jev call re-rank twenty pages at once?

The top candidates go to Jev together with per-page relevance, best-answer choice, and answerability questions, then blend into a new order. Pages outside the candidate pool cannot win.

Are the benchmark wins real or tuned to one docs site?

The reported corpus is 109 docs pages split into smaller units with 41 labeled queries. Jev search leads on intent hits with higher median latency, while the keyword pass alone matches older libraries.

What does the keyword pass do before Jev runs?

A weighted lexical scorer with prefixes, a light stemmer and one-edit typos runs in single-digit milliseconds, streams those hits immediately, and then Jev re-ranks the top 20 behind it.

How are Jev's relevance scores combined and cut?

Hits are re-ordered by 0.75 times per-page relevance plus 0.25 times best-answer share, anything below the 0.15 threshold is dropped, and the answer is cached in memory.

What happens if TypeSafe is slow or down?

The keyword order stands and the footer says so; timeoutMs defaults to 4000 before the component falls back to keyword ranking.

CommunityJev Integrations & Adapters

Jevsearch

Shadcn registry search block that shows lexical hits instantly then re-ranks the top candidates with Jev intent scoring in one request, with keyword fallback on failure.

See highlights

Open github.com

Stars: 8
Language: TypeScript
License: MIT
Published by: Community
Author: kylemclaren
Last verified: Oct 2026

kylemclaren/jevsearch

User repository on GitHub

View on GitHub

Site search that understands the question. Ranked by TypeSafe's Jev model.

Forks: 2
Last pushed: Updated Sep 2026

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

About this project

Site search that understands the question. A shadcn/ui registry block: a command palette that shows keyword hits on the first keystroke and, a couple of hundred milliseconds later, re-ranks them with TypeSafe's Jev model by what the visitor actually meant. Drop-in. One shadcn add against the registry JSON in this repo lands the component, hook, lexical index, server handler and an API route in your project. No package to wrap; the files are yours.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Ranking

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Streams keyword hits in single-digit milliseconds, then glides rows into Jev-judged order behind them.
Judges candidates with parallel relevance, best-answer, and answerability questions in a single request.
Needs no embeddings, vector store, or reindex job; a Markdown indexer builds the JSON index.
Falls back to keyword order with a notice after timeouts, and caches repeat queries in memory.

Watch out

Uncached queries cost input tokens per judged candidate set, so tune candidate counts and excerpt length for spend. 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

n8n-nodes-typesafe-ai

n8n community node for TypeSafe's System One API, exposing yes or no, choice and score questions as workflow nodes, with an Evaluate Questions operation that batches several over one state in a single request.