DocJev: open-source doc classifier splits PDFs in 210ms, 6x faster than GPT-5.6 Luna
llama_index · x · 2026-09-22
Jerry Liu introduced DocJev, an open-source library for document classification and splitting: feed it a document plus natural-language category rules, and it predicts document categories or sub-document boundaries.
- Speed: on a 40-PDF real-document benchmark, classification takes 139ms median vs 794ms for GPT-5.6 Luna (5.73x), splitting 210ms vs 1352ms (6.45x), with near-equal accuracy (40/40 classify; 7/8 split packets — Jev made one extra cut).
- Pluggable OCR backends: liteparse (fast, free, OSS) or LlamaParse (stronger for complex docs, higher latency).
- Total benchmark API cost was roughly 6.8 cents; an interactive visualization lets you inspect exactly where each engine cut, including a breakdown of the one boundary Jev got wrong.
Related event: LlamaIndex Founder Open-Sources DocJev for Faster Document Classification(3 posts)→
More from coding & agent
- Pi community ships 5 agent-team plugins as official sub-agents stay absent; Pi 0.87 splits session from context — solyarisoftware · 2026-09-22
- Pi v0.87.0 ships canonical session context editing, plus five breaking changes — solyarisoftware · 2026-09-22
- MiniMax details how to build a testbed for coding agent harness changes — MiniMax_AI · 2026-09-22
- JEV lands in game engine Spawn: AI bosses making 10 decisions per second — TAbrodi · 2026-09-22
- When Does Local AI Hardware Beat Cloud Models for Coding? Measure Cost per Accepted Change — Rama_Surasani_ · 2026-09-22
- Open-source Jev reads your WeChat and QQ chats to suggest replies, hits 1.8k GitHub stars — aigclink · 2026-09-22