GD4AI/obsidian-llm-wiki

What it solves

Finding connections between fragmented notes in a large Obsidian vault can be difficult. This plugin implements the "Karpathy LLM Wiki" concept, transforming a collection of flat notes into a structured, interconnected knowledge base (a wiki) where entities and concepts are automatically extracted and linked, making your personal knowledge graph queryable via AI.

How it works

The plugin reads your notes and uses a Large Language Model (LLM) to extract entities (people, organizations, products) and concepts (theories, methods). It then creates standalone wiki pages for these items in a dedicated wiki/ folder, complete with aliases and bidirectional [[wiki-links]].

Unlike traditional RAG (Retrieval-Augmented Generation) systems, it uses a zero-embedding, zero-vector-DB architecture. Instead, it employs Personalized PageRank (PPR) and Monte Carlo methods over the existing link graph to retrieve relevant information for queries, meaning it relies on the structure of your notes rather than mathematical vector embeddings.

Who it’s for

• Obsidian users who want to turn their notes into an AI-powered second brain without setting up complex external databases or Python environments.
• Privacy-conscious users who can run the system fully locally using Ollama or LM Studio.
• Knowledge workers who who want a queryable chat interface that answers based strictly on their own vault content.

Highlights

• Zero-Dependency Setup: A one-click Obsidian plugin with no requirement for embedding models, vector databases, or external runtimes.
• Broad Provider Support: Compatible with 16+ LLM providers including OpenAI, Anthropic, Gemini, Bedrock, and local options like Ollama.
• Multi-format Ingest: Supports Markdown, PDF, images, and Office files (via MinerU backend or native cloud provider PDF support).
• Graph-Based Retrieval: Uses Personalized PageRank (PPR) for retrieval instead of kNN/embeddings.
• Maintenance Tools: Includes a "Lint wiki" health scan and "Smart Fix All" to resolve duplicates, dead links, and orphans.