deeplightrag 1.0.22

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

DeepLightRAG

High-performance document indexing and retrieval system. Works with any LLM.

Features

• Dual-Layer Graph: Visual-Spatial + Entity-Relationship architecture
• GLiNER2 NER: Zero-shot entity extraction with fastino/gliner2-base-v1
• DeepSeek-OCR: Visual token compression for efficient document processing
• TOON Format: Token-efficient context formatting via toon-python
• Any LLM: Works with OpenAI, Gemini, Claude, Ollama, MLX

Installation

pip install deeplightrag

With GPU (CUDA):

pip install "deeplightrag[gpu]"

macOS (Apple Silicon):

pip install "deeplightrag[macos]"

Quick Start

Python API

from deeplightrag import DeepLightRAG

# Initialize
rag = DeepLightRAG()

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

# Retrieve
result = rag.retrieve("What are the key findings?")
print(result["context"])

CLI

# Index
deeplightrag index document.pdf

# Query
deeplightrag retrieve "What is the main topic?"

Architecture

┌─────────────────────────────────────────────┐
│              DeepLightRAG                   │
├─────────────────────────────────────────────┤
│  DeepSeek-OCR → Visual Token Compression    │
├─────────────────────────────────────────────┤
│  Dual-Layer Graph                           │
│  ├── Layer 1: Visual-Spatial (WHERE)        │
│  └── Layer 2: Entity-Relationship (WHAT)    │
├─────────────────────────────────────────────┤
│  GLiNER2 → Entity Extraction                │
├─────────────────────────────────────────────┤
│  Adaptive Retriever → Context Generation    │
└─────────────────────────────────────────────┘

Configuration

# config.yaml
ocr:
  model_name: "deepseek-ai/DeepSeek-OCR"
  resolution: "gundam"

ner:
  model_name: "fastino/gliner2-base-v1"
  confidence_threshold: 0.3

retrieval:
  top_k: 5

Requirements

• Python >= 3.9
• CUDA GPU recommended for DeepSeek-OCR
• Dependencies: gliner2, toon-python, networkx, torch

License

MIT License