Anji-Bridge (安济桥)

PDF to AI Agent Knowledge Bridge

Convert PDFs to enhanced, AI-agent-ready Markdown/JSON documents.

Features • Quick Start • Installation • Usage

---

English | 中文

---

What is Anji?

Anji bridges the gap between PDFs designed for human reading and the structured, semantic text that AI agents require. It leverages:

• PaddleOCR-VL for high-quality PDF-to-Markdown conversion
• Ovis2.5-9B Vision-Language Model for intelligent image analysis
• Mistune for flexible AST manipulation

Features

Feature	Description
Smart OCR	Extracts text, tables, and images with layout awareness
VLM Image Analysis	Generates captions and descriptions for embedded images
Decorative Filtering	Removes logos, watermarks, and noise automatically
Heading Correction(developing)	Fixes OCR-generated heading hierarchy issues
Multi-Format Output	Export to Markdown, JSON, or structured data
Batch Processing	Efficiently process multiple PDFs
Flexible Pipeline	Run full pipeline or individual steps
Base64 Embedding	Embed images as base64 data URLs in markdown

Quick Start

# Install
pip install -e .

# Convert a PDF
anji pipeline document.pdf output/

# Embed images as base64 (single portable file)
anji pipeline document.pdf output/ --embed-base64

# Or use as a Python library
python -c "
from anji import run_full_pipeline
run_full_pipeline('document.pdf', 'output/')
"

Installation

# Basic installation
pip install -e .

# With development dependencies
pip install -e ".[dev]"

Prerequisites

Anji requires two external services running:

1. PaddleOCR-VL Server (port 8118)

Requires GPU. Run using Docker:

docker run \
    -it \
    --rm \
    --gpus all \
    --network host \
    ccr-2vdh3abv-pub.cnc.bj.baidubce.com/paddlepaddle/paddleocr-genai-vllm-server:latest-nvidia-gpu \
    paddleocr genai_server --model_name PaddleOCR-VL-1.5-0.9B --host 0.0.0.0 --port 8118 --backend vllm

2. Ovis2.5-9B VLM Server (port 8000)

Requires GPU with ~16GB VRAM. Run with vLLM:

vllm serve AIDC-AI/Ovis2.5-9B \
    --trust-remote-code \
    --port 8000 \
    --gpu-memory-utilization 0.4

Note: If you encounter RuntimeError: Exception from the 'vlm' worker: only 0-dimensional arrays can be converted to Python scalars, install numpy==1.26.4.

Usage

Command Line

# Full pipeline
anji pipeline input.pdf output_dir

# Batch processing
anji batch output_base file1.pdf file2.pdf file3.pdf

# Individual steps
anji pdf input.pdf output_dir          # PDF → Markdown
anji image input.md output.md          # Analyze images
anji md enhance input.md output.md     # Enhance AST
anji md export input.md out --format json  # Export

Output Options

# Keep images folder (default: enabled)
anji pipeline input.pdf output/ --keep-images

# Disable images folder
anji pipeline input.pdf output/ --no-keep-images

# Embed images as base64 (single portable markdown file)
anji pipeline input.pdf output/ --embed-base64

# Combine options
anji pipeline input.pdf output/ --embed-base64 --no-keep-images

Python API

from anji import Pipeline, run_full_pipeline, batch_pipeline

# Simple usage
run_full_pipeline("document.pdf", "output/")

# Advanced usage
pipeline = Pipeline(
    paddleocr_server_url="http://localhost:8118/v1",
    vlm_server_url="http://localhost:8000/v1"
)

outputs = pipeline.run(
    input_path="document.pdf",
    output_folder="output",
    output_format="both",  # markdown, json, structured, or both
    keep_images=True,  # keep imgs folder
    embed_base64=False,  # or True for single file
)

# Batch processing
batch_pipeline(
    input_paths=["doc1.pdf", "doc2.pdf"],
    output_base_folder="batch_output"
)

pipeline.close()

Output Structure

output/
└── document_name/
    └── enhanced/
        ├── document.md     # Enhanced Markdown
        ├── document.json   # JSON AST (optional)
        └── imgs/          # Extracted images (optional)
            ├── image1.jpg
            └── image2.jpg

With --embed-base64, images are embedded directly in the markdown file as base64 data URLs.

How It Works

Anji processes PDFs through 4 stages:

1. PDF → Markdown - Uses PaddleOCR-VL to extract text, tables, and images
2. Markdown → AST - Parses markdown into an abstract syntax tree using Mistune
3. Enhance - Analyzes images with VLM, fixes heading levels, filters decorative elements
4. Export - Outputs as Markdown, JSON, or structured data

Configuration

Environment Variables

Variable	Default	Description
API_BASE_URL	http://localhost:8000/v1	VLM server URL
API_KEY	abc-123	VLM API key
MODEL_NAME	AIDC-AI/Ovis2.5-9B	VLM model name

CLI Options

anji pipeline input.pdf output/ \
  --format markdown|json|structured|both \
  --no-enhance \
  --no-fix-headings \
  --no-filter-decorative \
  --no-enrich-images \
  --keep-images \
  --embed-base64 \
  --dummy  # Test without API calls

Development

# Code formatting
black anji/

# Linting
ruff check anji/

# Type checking
mypy anji/

# Testing
pytest

Project Structure

anji/
├── anji/              # Main package
│   ├── __init__.py       # Exports
│   ├── main.py           # CLI entry point
│   ├── cli.py            # Command-line interface
│   ├── pipeline.py       # Pipeline orchestration
│   ├── pdf_converter.py  # PDF → Markdown
│   ├── image_analyzer.py # VLM image analysis
│   ├── ast_handler.py    # AST manipulation
│   ├── enhancement.py    # AST enhancement
│   └── exporters.py      # Export utilities
├── pyproject.toml        # Package configuration
├── README.md             # English documentation
├── README_CN.md          # Chinese documentation
├── CLAUDE.md             # Claude Code context
└── .gitignore

License

MIT License. See LICENSE for details.

Contributing

Contributions are welcome! Please read CLAUDE.md for development guidelines.

---

Built for AI agents, by AI agents