abhishekmamdapure/jev-information-extraction
User repository on GitHub
Parsing the PDF and extracting the relevant information
- Forks
- 2
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
Jev Information Extraction
FastAPI demo that turns PDF text into candidate answers to your questions with Jev, then maps the returned probabilities back to the exact chunks and page positions they came from.
- Stars
- 3
- Language
- Python
- Published by
- Community
- Author
- abhishekmamdapure
- Last verified
- Oct 2026
About this project
Ask questions about a PDF. Use Jev to rank the source text that answers them. Inspect each match, its probability, and its location on the original page. This interactive demo uses TypeSafe's jev-latest model through the Python SDK's systemone API. It turns extracted PDF text into candidate answers for natural-language questions such as ?What is the GST number?? or ?What is the total invoice amount??. A FastAPI service runs the evaluation and serves the Oat UI frontend from one URL.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Extraction stays local: pdf-inspector pulls embedded text chunks and positions, PDFium renders, and only then does Jev see the page.
- Each question becomes a Choice whose criteria map page-local chunk indexes to text, so one request per page covers every question.
- The UI shows the top chunk, an expandable runner-up, probability meters keyed at 80% and 50%, and red boxes over the rendered page.
- Up to eight page tasks run in parallel, and the API returns the resolved model name and token usage for each successful page.
Watch out
The README is explicit that this is question-to-text-chunk selection, not a benchmark: nothing computes accuracy, precision, recall or F1, and a displayed probability is the model's score for one candidate, not proof the answer is right. Source text is shown exactly as extracted, so chunks are never stitched into a rewritten answer and values are not normalized — asking for only the number can still return a currency label. There is no OCR, password-protected PDFs are unsupported, and the limits are 20 MB and 200 pages. There is no authentication and no per-user isolation, and documents live in process memory until restart, which the README calls a demo rather than a private multi-user document service and which is why the Railway config pins one replica. 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.
No. Evaluation here means question-to-text-chunk selection, performed independently per selected page, and the app calculates no accuracy, precision, recall or F1. A meter is the model's score for one candidate chunk. The minimum chunk probability setting, default 0.9, is only the highlight cutoff — it does not change the model request and does not hide ranked answers.
No. Extraction works on embedded text, so a value broken across chunks is not automatically recombined, and the boxes mark extracted text regions rather than semantic fields. If you ask for only the number, expect a currency label in the matched chunk to survive, because the app shows source text as extracted instead of rewriting it.
Upload, extraction and preview work with no key at all. The key lives in the page's in-memory field, not browser storage, and goes to the backend per evaluation for a request-scoped client; it is never written to a file or an environment variable. For evaluation, page text, questions and candidate text are sent to TypeSafe — the model call does not send the PDF image. PDFs are held in server process memory until restart, with no auth, no per-user isolation and no automatic expiry.
The README documents a single-service Railway deploy with a Python 3.12 Dockerfile, railway.json and /api/health, deliberately at one replica because documents are stored in process memory. There are no app environment variables; users paste their own session key in the UI. Anything multi-tenant or restart-tolerant is work the README does not cover, and it is explicit that this is a demo.
Each question becomes a Choice whose instructions hold the question and whose criteria map page-local chunk indexes to the extracted text. All questions for a page go in a single system_one call with model jev-latest, up to eight page tasks run in parallel, and the returned probabilities are mapped back to the original chunks. Pages without embedded text are skipped, and there is no cross-page aggregation.
PDFs must be 20 MB or smaller and between 1 and 200 pages; password-protected files are unsupported. OCR is not enabled, so a scanned page with no text layer cannot be evaluated and is skipped. The first-N-pages choice limits model evaluation, not the initial extraction, and uploading alone makes no TypeSafe request.
The resolved model name and token usage for each successful page evaluation, plus the UI's total batch wall time and per-page render and evaluation timings. Batch render timing includes any wait for the rendering lock and is not the duration of every later preview refresh, so the two numbers answer different questions.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPython164GitHub stars
End-to-end toolkit that synthesizes decision data, trains and calibrates candidate scorers, evaluates fixed splits, and serves Jev-compatible inference for interactive web decisions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zwliJay/jev-forgePython
zwliJay/jev-forgePython63GitHub stars
Claude Code kit that guards PreToolUse calls with a free pre-filter plus one Jev question for ambiguous cases, and adds sub-agent sizing, rechecks, search, daemon speedups, and monitoring.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
jonathanavis96/jev-kitPython
jonathanavis96/jev-kitPython70GitHub stars
Self-hosted proxy that records typed decisions, distils them into small heads on a frozen encoder, and answers locally when confident.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bladedevoff/stuntdPython
bladedevoff/stuntdPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.