jerryjliu/docjev

What it solves

DocJev provides a way to automatically classify documents and split large PDF packets into individual documents based on natural-language rules. It solves the problem of manually sorting through multi-page files (like a combined PDF of various financial reports) and identifying where one document ends and another begins, even when adjacent documents belong to the same category.

How it works

The system uses a pipeline of OCR and a decision engine. It extracts text from PDF, DOCX, or PPTX files using LiteParse (locally) or LlamaParse (cloud-based OCR for complex layouts). This text is then sent to the Jev decision engine (a hosted service) which applies natural-language category rules to predict the document's category or the boundaries between pages. The process is page-level and contiguous, meaning it identifies page ranges for each segment.

Who it’s for

It is designed for users and developers who need to process large volumes of unstructured document packets, particularly in finance, government, or legal sectors where documents like tax forms, press releases, and financial reports are often bundled together.

Highlights

• Multi-format support: Handles PDF, DOCX, and PPTX files.
• Flexible rules: Uses YAML-based natural-language descriptions to define categories and splitting logic.
• Hybrid OCR: Offers local parsing via LiteParse and optional cloud-based OCR via LlamaParse.
• Review flags: Automatically flags uncertain boundaries or category conflicts for human review.
• Performance benchmarks: Includes detailed comparisons of decision latency and accuracy against other models like GPT-5.6 Luna.