Published
DocJev classifies and splits real PDFs with Jev, 5.7x to 6.5x faster than Luna on the author's set
Jerry Liu posted DocJev, an Apache 2.0 library that classifies and splits PDFs with local LiteParse and hosted Jev. On 40 public-sector originals, both engines scored 40/40. Jev split 7 of 8 packets exactly against Luna 8/8, at 5.73x and 6.45x the median decision speed.
Jerry Liu posted DocJev on September 20. He is cofounder and CEO of LlamaIndex. The repo is jerryjliu/docjev, Apache 2.0. The tweet said it is 6x faster than gpt-5.6-luna with equivalent accuracy.
You give the library a PDF, DOCX, or PPTX and a YAML of natural-language category rules. LiteParse extracts complete page text on the machine. Hosted Jev then classifies the document or marks the page ranges that split a packet into its component publications. LlamaParse is an optional cloud OCR path. Local OCR does not make the decision offline. The README says the project does not call LlamaIndex’s hosted Classify or Split APIs.
The 40-document accuracy pilot is in the repo as real-small-v1-run01. Forty authentic English public-sector PDFs, eight per category, reused across eight constructed five-document packets, 116 unique pages. Both engines saw the same LiteParse text. Decision times exclude OCR. Classification: both 40/40. Jev median 138.6 ms, Luna 794.3 ms, 5.73x. Splitting: Jev 7/8 exact packets, Luna 8/8. Medians 209.6 ms and 1,352.3 ms, 6.45x. Both found all 32 true boundaries, including four adjacent same-category Treasury-style pairs, and labeled every packet page. Jev added one extra cut before a Federal Reserve statement’s implementation attachment; the frozen rules keep that attachment with the publication.
Estimated decision cost on that pass was $0.011663 for Jev and $0.046894 for Luna. Human annotation review was not performed. The README treats the set as a small convenience sample and does not claim a confidence interval.
A 15-page public-finance demo packet (IRS Form 941, two Treasury auction results, a BEA release) is in examples/real. The two adjacent Treasury pages share a category and still split. A September 21 follow-up from Liu points at the visual report in the README.
We did not install the package.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Jerry Liu, LlamaIndex cofounder and CEO, posted DocJev on September 20. The tweet claims 6x faster than gpt-5.6-luna with equivalent accuracy. The README is the spec: Apache 2.0, Python 3.11+, CLI docjev. LiteParse extracts page text locally; Jev is a hosted TypeSafe call. LlamaParse is optional OCR. The library does not call LlamaIndex's hosted Classify or Split APIs. The 40-document accuracy pilot (real-small-v1-run01) used 40 authentic PDFs, eight per category, reused across eight five-document packets, 116 unique pages. Both engines received the same LiteParse text. Decision times exclude OCR. Classification: both 40/40, Jev median 138.6 ms, Luna 794.3 ms, ratio 5.73x. Splitting: Jev 7/8 exact packets, Luna 8/8, medians 209.6 ms and 1,352.3 ms, ratio 6.45x. Both found all 32 true boundaries, including four adjacent same-category ones, and labeled all 116 packet pages. Jev added one extra boundary before a Federal Reserve statement's implementation attachment. Estimated decision cost $0.011663 for Jev and $0.046894 for Luna. Human annotation review was not performed. The README calls this a small convenience sample. We did not run the library. TypeSafe's Master Customer Agreement section 2.3(f) forbids publishing benchmarks of the Services; the figures are reported as published.
Compare
Bryo and Vogel sort email. Every scored 37 writing samples. DocJev is page-level classify and split on PDFs, with a Luna baseline on the same LiteParse text. The tweet's "equivalent accuracy" is the 40/40 classify tie; Luna won the 8-packet split 8/8 to 7/8. JevBench still has Luna ahead on hard-tier accuracy (94.5% versus 74.1%) and behind on the composite. Richard Tang posted JevGraph on top of DocJev the same day.
Terms
- DocJev
- Jerry Liu's Apache 2.0 library for classifying and splitting PDF, DOCX, and PPTX files. LiteParse reads pages locally; hosted Jev returns the category or the page-range boundaries.
- LiteParse
- The default local parser in DocJev. It extracts complete page text on the machine. Decision timing in the published pilots excludes this step.