Why not just build a vector index over the PDF?

Because Jev turns ranking into a typed number per line. Each line becomes one noul question, does this line answer the query, and the app ranks by the returned probability, so there is no prompt parsing, no answer text to hallucinate, and no chunking strategy to tune. Extract the text once and keep asking, with thresholds and question wording in one config file.

What stops someone abusing the deployed proxy's API key?

The server-side proxy accepts only same-origin POSTs from the app, pins the model, caps request size and rate-limits each IP. In development the key is attached by a Vite middleware serving /api/jev rather than handed to the browser, and .env.local is gitignored. The live proxy spends real credits, so it is deliberately the narrowest path.

Will semantic search find me the sentence I only half remember?

Not reliably. The unit of retrieval is one extracted text line, so a claim spread across a sentence boundary, a footnote or a table row may fall below the hit threshold. When nothing clears the threshold the app shows the closest lines instead, which is the honest signal that the query did not match rather than a confident wrong answer.

What does one question actually cost, and how many calls is that?

Sixteen lines per request with sixteen requests in flight, so a document costs roughly one call per sixteen lines however many questions you ask, since all questions in a request share the same page text. At $0.042 per million input tokens with free output, the README puts every line of a 15-page paper well under a cent, and 429, 529 and 5xx responses get exponential backoff with jitter.

What question does JevPDF ask for each line?

A noul question, does this line answer the query, one per line, returning a probability from 0 to 1 that the app uses to rank the highlights.

How does JevPDF keep its API key off the client?

The key stays on the server; /api/jev is a Vite middleware that adds the Bearer header in dev and preview, and the deployed Bun proxy only accepts same-origin POSTs, pins the model, caps request size and rate-limits each IP.

How does JevPDF extract and cache a PDF?

pdf.js reads each page's text runs and groups them into lines with rects on the client, cached in IndexedDB by the file's SHA-256; Jev only receives text, never PDF bytes.

CommunityJev Tools & CLIs

Jevpdf

PDF search that asks Jev one noul per line, does this line answer the query, and ranks the lines by the returned probability, so matches highlight without embeddings, a vector index or a chunking strategy. Text is extracted locally with pdf.js and the key stays on the server.

See highlights

Open github.com

Stars: 7
Language: TypeScript
License: MIT
Published by: Community
Author: kylemclaren
Last verified: Oct 2026

kylemclaren/jevpdf

User repository on GitHub

View on GitHub

Ask a PDF in your own words and watch the matching lines light up. React + pdf.js + TypeSafe Jev.

Forks: 2
Last pushed: Updated Sep 2026

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

About this project

Ctrl+F for when you don't know the exact words. Ask a PDF in your own words and watch the matching lines light up. Run · Using it · How it works · Deploy. JevPDF is built on Jev, TypeSafe's flagship System One model. Jev doesn't write prose. It answers typed questions with structured values your code can use directly, and that's exactly what search needs.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

One narrow noul per line, with all questions in a request scored in parallel against the same shared page text.
It sends 16 lines per request and keeps 16 requests in flight, so highlights stream in page by page as batches return.
At $0.042 per million input tokens with free output, checking every line of a 15-page paper costs well under a cent.
Exact-text mode is local and case- and whitespace-insensitive, can cross line breaks, and nouls are cached per file hash and query.

Watch out

Cost is low per query but scales with the number of lines rather than with the number of questions you type, so repeated reformulations on a long document keep adding up, and caching by file hash and query only helps when you repeat yourself. The retrieval unit is a single extracted line, so an answer spread across a sentence boundary, a footnote or a table row can score below the hit threshold, and nothing outside the page text is ever consulted. Requests are checked against the 64k and 32k token limits, which shapes what fits on a dense page. 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

jev-recall

Python memory filter that judges every stored memory against the current request with one calibrated yes/no per memory in a single Jev request, keeping everything above a threshold.