This project transforms raw Markdown notes into a self-improving, interlinked personal wiki using local LLMs via Ollama. It enables users to build a zero-sharing, locally-controlled knowledge base that automatically synthesizes and updates information, integrating seamlessly with Obsidian.

How It Works

A three-stage pipeline orchestrates LLM usage: ingest extracts concepts from notes using a fast LLM; compile generates wiki articles with wikilinks via a heavy LLM, incorporating user feedback from rejected drafts; review provides an interactive interface for draft approval. This local-first, provider-flexible approach avoids vector databases, relying on an index-based query system and Obsidian's native features. User data remains private, and the wiki compounds knowledge over time.

Quick Start & Requirements

• Installation: pip install obsidian-llm-wiki or uv tool install obsidian-llm-wiki.
• Prerequisites: Python, Ollama (or compatible endpoint), LLM models (e.g., gemma4:e4b for fast analysis, qwen2.5:14b for heavy writing), Obsidian.
• Setup: Install package, set up Ollama and pull models, run olw setup wizard, initialize vault (olw init), add notes to raw/, run olw run.
• Links: Ollama Download: https://ollama.com/download.

Highlighted Details

• Concept Extraction & Incremental Updates: Each concept gets an article, updated based on source changes.
• Rejection Feedback Loop: User feedback on rejected drafts refines future LLM outputs.
• Multi-Provider & Multi-Language: Supports various LLM endpoints and auto-detects/generates articles in note languages.
• olw compare: Safely previews model/provider switches in isolated vaults.
• Offline Test Suite: Comprehensive tests available without external LLM dependencies.

Maintenance & Community

Maintained by kytmanov. No specific community channels or external sponsorships detailed.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permissive for commercial use and integration with closed-source projects.

Limitations & Caveats

The index-based query system is optimized for ~100 source notes. Manual draft review is default. LLM output quality depends on model choice and hardware. Concepts blocked after rejections require manual unblocking.