OCR and Parsing Pipelines

Parse Engine

Mode	Meaning	Best for
local_ocr	Parse PDF locally, then run OCR as needed	General default mode
remote_ocr	Remote OCR-first pipeline	OCR-quality-first scenarios
baidu_doc	Baidu document parsing pipeline	Structured parsing
mineru_cloud	MinerU cloud parsing pipeline	Tables, formulas, and complex layouts

OCR Provider

Provider	Meaning	Notes
aiocr	Remote OpenAI-compatible OCR	Good for higher-quality OCR
tesseract	Local Tesseract	Fewer external dependencies
paddle_local	Local PaddleOCR	Fully local setup
baidu	Baidu OCR	Standalone provider

AIOCR Chain

Mode	Meaning	Characteristic
direct	Send the whole page directly to a vision model	Simplest setup
layout_block	Split into local blocks, then recognize block by block	Better for dense mixed layouts
doc_parser	Structured document parsing channel	Better structural information

Scanned Page Mode

Mode	Meaning	Result
fullpage	Keep the page as a background and overlay editable text	Safest and closest to the original
segmented	Split charts, screenshots, and other image regions into separate objects when possible	More editable afterwards

Recommended Starting Configuration

For the highest chance of a stable first run, start with:

• remote_ocr
• aiocr
• fullpage

Then adjust based on your needs:

• try segmented when you want more editable image regions
• try baidu_doc or mineru_cloud when structural parsing matters more

Boundaries

This project is better understood as a high-fidelity reconstruction tool for scanned and image-heavy documents. It should not be interpreted as:

• a guarantee that any PDF becomes a fully structured, fully editable native PPT
• a promise that complex documents work well without OCR or parsing configuration
• a system where every page becomes easier to edit than the original source