The development of Large Language Models (LLMs) has introduced powerful capabilities for information processing. However, the effectiveness of these models is intrinsically linked to the quality and accessibility of the data they utilize. This article outlines the construction of an "LLM Wiki," a structured system designed to transform raw information into a persistent, accessible knowledge base for both human users and AI agents. This system aims to create a shared memory layer that evolves and improves with the addition of new information, aligning with individual thinking patterns.

The Three-Layered Architecture

The LLM Wiki system is built upon three fundamental layers:

1. Sources (Raw Layer): This layer comprises the captured source material, such as transcripts, documents, or articles. This raw data serves as the input for the Wiki.
2. Wiki (Knowledge Layer): This layer is where topics, concepts, entities, and projects are extracted from the raw source material. The Wiki is organized and maintained according to a defined schema.
3. Schema (Organizational Layer): This layer dictates how the Wiki is structured and maintained. It includes elements like agent configurations, templates, skills, and command documentation, which instruct AI agents on how to manage the Wiki.

When new information is added to the Wiki, it cites the original source material, ensuring the preservation of the raw data while creating a structured, interconnected knowledge graph. The schema acts as a contract between the user and the AI agent, ensuring consistency and alignment.

Demonstration of the Workflow

A practical demonstration illustrates the functionality of these layers. A YouTube video transcript on "attention and how to fix it" was used as a source.

1. Ingestion: The transcript was captured using an Obsidian web clipper and saved to the "raw" folder within the Obsidian vault.
2. Extraction: An AI agent was instructed to check for new raw files, ingest them into the Wiki, and run maintenance scripts. The agent identified the new source and proceeded to extract concepts.
3. Maintenance: The system then entered a maintenance or "linting" mode, verifying the accuracy and consistency of the newly added information and ensuring adherence to the defined schema.

The Obsidian graph view visually represents these layers, with "raw" material in green, "Wiki" content in blue, and the "schema" in orange. This visualization highlights the interconnectedness of the data as new sources are added and concepts are extracted.

Prerequisites for Building the LLM Wiki

To establish the LLM Wiki system, several tools and software are required:

• Python: Essential for running repeatable scripts, enhancing accuracy, and saving on token usage.
• Git: Provides version control, enabling the creation of save points and facilitating recovery from potential errors made by AI agents.
• Obsidian: Serves as the core memory system and the second brain layer for storing files and notes.
• Agentic AI (Coding Agent): The central component that powers the system, running scripts, processing files, and suggesting improvements. The author uses Codex but notes that any agentic AI can be utilized.
• Optional: Local Model: For free and private operation, a local LLM can be integrated.

Core Setup and Workflow

The construction of the LLM Wiki involves the following steps:

1. Initialize Obsidian Vault as a Git Repository: Create a new Obsidian vault and initialize it as a Git repository to enable version control. A .gitignore file is used to exclude unnecessary configuration files.
2. Configure Agentic Environment: Ensure the chosen agentic environment has Obsidian-aware skills. This involves installing specific plugins or skill packs that allow the agent to interact effectively with Obsidian's wiki-linking system. The author recommends the Obsidian CLI, markdown basis, canvas, and defuddle skills, and emphasizes enabling the command-line interface in Obsidian settings.
3. Create Folder Structure: Establish a core folder structure within the Obsidian vault to organize different types of information, including "raw," "wiki," and "schema" folders.
4. Implement Schema Files: Introduce schema files, such as agents.md, which act as a constitution for the AI agent, defining rules, workflow guidance, and structural requirements. This includes specifying front matter schemas, naming conventions, and workflow examples to ensure repeatable behavior.
5. Develop Templates: Create templates for various note types (e.g., sources, concepts, topics, entities, projects, logs) to ensure consistent formatting and property inclusion. These templates are linked to the front matter schema.
6. Integrate Python Scripts (Tooling): Develop Python scripts that provide the agent with specific commands for tasks like building an index, checking structure, querying, and performing maintenance. These scripts translate schema rules into executable code for consistent operation.
7. Ingest and Validate Raw Data: Test the system by ingesting a new raw source file (e.g., a video transcript). The agent uses the ingest skill to process the file, extract concepts, and update the Wiki.
8. Run Maintenance and Linting: Execute maintenance scripts to validate the ingested data, update the catalog, and ensure adherence to all defined rules and schema requirements. This includes checking for routing gaps or catalog search failures.

Advanced Features

Beyond the core setup, several advanced features can enhance the LLM Wiki system:

• Obsidian Safe System (Firewall): Implement a wrapper around the Obsidian CLI skill to restrict agent access to specific, approved vaults, preventing unauthorized modifications to other vaults.
• Local LLM Setup: Integrate local LLM models (e.g., via Ollama) for private and free operation. This may involve creating a review folder for proposed changes before applying them to the main system, acting as a "human in the loop."
• Launcher Commands: Develop one-click commands (e.g., for Mac or Windows) to initiate local LLM workflows, simplifying the process of updating the Obsidian vault.
• PDF Ingestion: Develop a system to extract markdown from PDF files and integrate them into the LLM Wiki.
• Molecular Zettelkasten System: Implement advanced knowledge theory for a more sophisticated knowledge organization beyond a standard wiki.
• Querying with RAG and Graph RAG: Enhance the querying capabilities using Retrieval Augmented Generation (RAG) and Graph RAG for more intelligent information retrieval.

The LLM Wiki is presented as a foundational system that can be continuously iterated upon and customized to align with individual workflows and thinking patterns. The author emphasizes building a solid foundation before integrating more complex features like local models, which may initially have lower intelligence compared to cloud-based systems. The ultimate goal is to create a dynamic, evolving knowledge base that improves the quality, connections, and efficiency of AI agents while reducing hallucinations.