VLM4OCR

vlm4ocr is a toolkit for Optical character recognition (OCR) with Vision language models (VLMs). In includes three components:

• Web Application for drag-and-drop access
• CLI for command line access
• Python package for Python access

What's new in v0.6.0

• OCR pipelines — process each page of a heterogeneous document differently. IndependentPagePipeline turns a per-page function into a concurrent, files-in / OCRResult-out pipeline; OCREngine.ocr_image_async is the atomic single-image call it builds on. Ideal for classifying and routing mixed form types in one PDF to different prompts/schemas. See OCR Pipelines.
• OCRPage.metadata — each page carries a free-form metadata dict (e.g. a page type assigned by a routing pipeline), preserved on the resulting OCRResult.

What's new in v0.5.0

• BBox output mode — OCREngine(output_mode="bbox", ...) returns text with bounding-box coordinates and labels. Leave user_prompt empty for full-text bbox OCR or set it to a free-text instruction (e.g., "patient name and DOB") for targeted extraction. Built-in format registry covers Qwen3-VL, Gemma 3/4, and GPT-4.1. See Quick Start.
• Web app — BBox tab — visualize bounding boxes directly in the browser with an Image | Raw response toggle. Batch mode emits annotated PNGs and a consolidated JSON per file.
• OCRPage dataclass — OCRResult.pages entries are now OCRPage dataclasses with .text, .bboxes, .image_width, .image_height, and a .plot_bboxes() helper. Dict-style access still works.