deeplightrag 1.0.15

DeepLightRAG: High-performance Document Indexing and Retrieval System (use with any LLM)

DeepLightRAG

DeepLightRAG is a high-performance document indexing and retrieval system designed to work with any Large Language Model (LLM). It features a dual-layer graph architecture (Visual-Spatial and Entity-Relationship) to provide context-aware and visually-grounded retrieval.

Features

• Dual-Layer Graph: Combines visual layout awareness with semantic entity relationships.
• Visual-Grounded Retrieval: Retrieves not just text, but visual regions and their spatial context.
• Robust OCR: Integrated with DeepSeek-OCR and EasyOCR fallback for reliable text extraction.
• Advanced NER: Uses GLiNER for zero-shot entity recognition.
• Flexible LLM Support: Compatible with OpenAI, Google Gemini, Anthropic, and local LLMs via MLX/Ollama.

Installation

Standard Installation

pip install deeplightrag

With GPU Support (NVIDIA CUDA)

For optimized performance using quantization (4-bit/8-bit):

pip install "deeplightrag[gpu]"

For macOS (Apple Silicon)

For optimization on M1/M2/M3 chips:

pip install "deeplightrag[macos]"

Usage

Command Line Interface

Index a document:

# Basic usage
deeplightrag index document.pdf

# With custom configuration
deeplightrag index document.pdf --config config.yaml

Retrieve information:

deeplightrag retrieve "What is the main topic?" --config config.yaml

Configuration File (config.yaml)

You can customize the model and system behavior using a YAML file:

ocr:
  model_name: "deepseek-ai/deepseek-ocr"
  # Override MLX automatic selection (useful for some models)
  use_mlx: false 
  resolution: "base"

retrieval:
  top_k: 5
  rerank: true

Python API

from deeplightrag.core import DeepLightRAG

# Initialize with hardware auto-detection
rag = DeepLightRAG(config={"ocr": {"use_mlx": True}})

# Index
rag.index_document("research_paper.pdf", document_id="doc_001")

# Retrieve
result = rag.retrieve("Summarize the methodology")
print(result)

License

MIT License