LLAutoLibrary

LLAutoLibrary is an automated personal knowledge-graph system designed for researchers. It transforms raw, unstructured documents into a structured, interconnected “Digital Garden” or wiki. By leveraging local LLMs and vector databases, LLAutoLibrary automatically extracts core concepts from your documents and links them together, allowing for efficient discovery and high-fidelity information retrieval.

Inspired by Andrej Karpathy’s vision of personal AI operating systems and the potential of LLMs as a reasoning layer over structured data, this project aims to bridge the gap between static document storage and dynamic knowledge management.

The Vision

The goal of LLAutoLibrary is to move beyond simple folder structures. In a professional or organizational setting, this system allows an LLM to sit atop a massive knowledge graph, enabling it to generate insights, summaries, and reports based on the explicit relationships between data points rather than just raw text chunks.

Planned Additions

• Linting to clean up overlapping concepts, missing information, and standardizing file formats
• Ability to toggle between local and cloud models
• Containerization
• Additional tools (data visualizations, presentations, reports)
• Improved UI/UX

Technology Stack

This project is built with a focus on local privacy, high performance, and modern web standards:

• LLM Engine: Ollama running Gemma 4 e4b (or preferred models) locally for privacy-preserving document processing and concept extraction.

• Vector Database: LanceDB used for high-performance vector search to identify and deduplicate concepts across the taxonomy.

• Backend: Python-based engine (engine.py) for document parsing, metadata extraction, and graph construction.

• Frontend: A modern, responsive interface built with React, TypeScript, and Vite, featuring a specialized Research View and Chat interface.

• Data Format: Interoperable Markdown files, structured to be compatible with tools like Obsidian for local-first knowledge management.

Features

• Automated Parsing: Drop raw PDFs or documents into the /raw folder and let the engine convert them into structured Markdown.

• Concept Extraction: The LLM automatically identifies key themes and creates dedicated “Concept” pages.

• Semantic Linking: Uses vector similarity to link new documents to existing concepts in your library.

• Research View: A dedicated UI for exploring the connections between your sources and extracted knowledge.

Installation

If you just want to test the LLM functionality, use testing.ipynb

Follow these steps to set up LLAutoLibrary on your local machine.

Prerequisites

• Python 3.10+
• Node.js (v18+) & npm
• Ollama: Download and install Ollama. Once installed, pull the model:

ollama pull gemma4:e4b

• Gemma 4: This version requires at least 16 GB of VRAM, or 16 GB of RAM on Apple Silicon. If you want to downgrade, you can use gemma4:e2b instead.
  1. Clone the Repository

     git clone https://github.com/etcasketera/LLAutoLibrary.git
     cd LLAutoLibrary
     

  2. Backend Setup (Engine): Create a virtual environment and install the required Python packages (I used conda)

Windows

python -m venv venv
.\venv\Scripts\activate

Mac/Linux

python3 -m venv venv
source venv/bin/activate

pip install -r requirements.txt

1. Frontend Setup Navigate to the frontend directory and install the dependencies:

cd frontend
npm install

1. Running the Project Start the Engine: Run the Python script to begin processing documents in your /raw folder.

python main.py

Start the UI: In a new terminal, run the development server:

cd frontend
npm run dev

Open your browser to the local URL provided by Vite (usually http://localhost:5173).

Acknowledgements & Inspiration

• Andrej Karpathy: For the inspiration regarding the future of personal AI and the “Software 2.0” stack.

• Open-Source Tools: This project is made possible by the incredible communities behind Ollama, LanceDB, React, and the Gemma model family by Google.