This package parses PDF documents by identifying distinct content regions (text, figures, tables, etc.) using a layout analysis model. It then feeds images of these regions to multimodal LLMs like GPT-4o or Qwen-VL to extract structured text, making it suitable for RAG applications.

How It Works

The core approach leverages a layout analysis model to segment each PDF page into categorized regions, including titles, text, figures, captions, tables, headers, footers, references, and equations. Each region is assigned coordinates and a reading order. Images of these identified regions are then processed by multimodal LLMs, enabling more precise content extraction compared to traditional text-only PDF parsers.

Quick Start & Requirements

• Install via pip: pip install llmdocparser
• Requires Poetry for installation from source.
• Usage involves providing a PDF path, output directory, and LLM configuration (e.g., Azure or OpenAI API keys, endpoint, deployment).

Highlighted Details

• Identifies 10 distinct content region types within PDFs.
• Supports multiple LLM providers: Azure, OpenAI, and DashScope.
• Provides cost estimation for LLM processing, with an example showing $0.32 for a 15-page document using GPT-4o.
• Extracts content into a structured format suitable for RAG pipelines.

Maintenance & Community

• Project is hosted on GitHub: https://github.com/lazyFrogLOL/llmdocparser
• Star history is available.

Licensing & Compatibility

• The README does not explicitly state a license.

Limitations & Caveats

• The effectiveness of the layout analysis model and LLM extraction is dependent on the quality and complexity of the input PDFs.
• Specific LLM API keys and configurations are required for operation.