🧠 LLUMINAI — Advanced LLM & Agentic AI Research Lab

Comprehensive, Reproducible Research Playground for LLM Pretraining, Knowledge Graph RAG, Multi-Agent LangGraph Architectures, & ColPali Visual Retrieval.

Overview • Research Modules • Code Architecture • System Pipeline • Structure • Quick Start • License

---

📌 Overview

LLUMINAI is an extensive, reproducible experimental laboratory containing 20+ specialized modules exploring the frontier of Large Language Models and Autonomous AI Agents. Covering everything from scratch transformer pretraining and scaling law dynamics to stateful LangGraph multi-agent coordination, GraphRAG knowledge structures, ColPali vision-language multi-vector retrieval, and quantitative RAGAS evaluation.

---

✨ Research Modules & Capabilities

Module	Primary Focus	Outcome & Real Proof
🚀 LLM Pretraining & Scaling	Modules 11–13, 16	Scratch tokenization, causal transformer pretraining loops, and empirical compute scaling analysis
🕸️ Knowledge Graph & Advanced RAG	Modules 25, 34, 42, 43	GraphRAG entity-relationship extraction, Memo-RAG memory hierarchies, and cache-optimized vector search
👁️ ColPali Visual Document RAG	Module 34	Multi-vector patch-level visual embedding for dense charts, tables, and complex PDF layouts
🤖 Multi-Agent Orchestration	Modules 5, 6, 30, 36	Stateful LangGraph multi-agent cyclic graphs with SQLite checkpoints, human-in-the-loop, and tool routing
🎬 Multimodal VLM & Video Analysis	Modules 27, 53	Vision-language keyframe reasoning, automated semantic video cutting, and narrative arc deconstruction
⚖️ RAGAS Quantitative Evaluation	Modules 28, 29, 48	Automated evaluation of faithfulness, answer relevancy, context recall, and reinforcement fine-tuning logs

---

🔬 Code Architecture & Implementation

📐 Technical Highlights Across Core Modules

• 5_2025_11_21_langgraph/main.py: Builds stateful StateGraph(AgentState) workflows connecting research agents, tool execution nodes (ToolNode), and conditional router edges with SQLite-backed thread checkpoints (SqliteSaver).
• 34_2026_1_28_colpali_rag/: Implements vision-based multi-vector document retrieval using ColPali / Byaldi architectures, indexing raw PDF page image tokens directly without brittle OCR text extraction.
• 25_2026_1_7_graphrag/: Extracts entity-relation triplets using structured LLM schemas and builds query-time sub-graph communities for global semantic synthesis.
• 53_2026_3_11_vlm_video_cut/: Combines OpenCV frame difference metrics with Vision-Language Models to detect scene boundaries and generate timestamped editing EDL cuts.
• Container Infrastructure (Dockerfile, docker-compose.yml): Unified PyTorch + CUDA development stack with mounted live volume mapping and automated environment bootstrapping.

---

📊 System Pipeline

graph TD
  subgraph Ingestion [📚 Multimodal Ingestion Layer]
    Doc[📄 Complex PDFs / Excel] --> ColPali[👁️ ColPali Visual Vectorizer]
    Text[📝 Raw Texts & Transcripts] --> Graph[🕸️ GraphRAG Entity Builder]
    Video[📹 Video MP4] --> VLM[🎬 VLM Semantic Scene Cutter]
  end

  subgraph Reasoning [🧠 Multi-Agent LangGraph Core]
    ColPali & Graph & VLM --> Storage[(🗄️ Hybrid Vector & Knowledge Graph DB)]
    Storage --> Orchestrator[🚪 LangGraph StateGraph Orchestrator]
    Orchestrator --> AgentA[🤖 Researcher Agent]
    Orchestrator --> AgentB[✍️ Synthesis & Code Agent]
  end

  subgraph Validation [⚖️ Quantitative Benchmark]
    AgentA & AgentB --> RAGAS[📊 RAGAS Evaluation Engine]
    RAGAS --> Output((🏆 Verified Scientific Output))
  end

  classDef primary fill:#a855f7,stroke:#7c3aed,stroke-width:2px,color:#fff;
  classDef accent fill:#38bdf8,stroke:#0284c7,stroke-width:2px,color:#fff;
  classDef gold fill:#f59e0b,stroke:#d97706,stroke-width:2px,color:#fff;
  class Ingestion,Storage primary;
  class Orchestrator,AgentA,AgentB accent;
  class RAGAS,Output gold;

Loading

---

📁 Repository Directory Layout

lluminai/
├── 📁 assets/                 # High-resolution SVG banners
│   └── 🎨 hero.svg
├── 📁 1_2025_11_7_deep_ideation...    # Autonomous research idea generator
├── 📁 5_2025_11_21_langgraph...       # Stateful multi-agent graph workflows
├── 📁 11_2025_12_5_llm_pretraining_1..# Scratch causal transformer training
├── 📁 25_2026_1_7_graphrag...         # Knowledge graph RAG implementation
├── 📁 34_2026_1_28_colpali_rag...     # Multi-vector vision-language RAG
├── 📁 53_2026_3_11_vlm_video_cut...   # Vision-language video segmentation
├── 🐳 Dockerfile & docker-compose.yml # Containerized GPU environment
└── 📄 README.md                       # Complete research documentation

---

🚀 Quick Start

1. Docker Compose (GPU Accelerated)

# 1. Start GPU-accelerated workspace
docker compose up -d

# 2. Access JupyterLab environment at http://localhost:8888

2. Standalone Module Execution

# Navigate to specific experimental module
cd 5_2025_11_21_langgraph
pip install -r requirements.txt
python main.py

---

Released under the MIT License. Crafted with ❤️ by LoNebula