llm-wiki

中文用户建议直接阅读：README.zh-CN.md

For English readers, this page remains the default project homepage.

A local-first, continuously evolving knowledge base built around Karpathy’s LLM Wiki pattern.

Instead of treating documents as files to search over and re-interpret every time, this project treats ingestion as compilation: each new source is transformed into structured wiki nodes that can accumulate, interlink, and improve over time.

At a Glance

• What it is: a local-first LLM Wiki workflow
• Core runtime: LM Studio + Docker + Python watchdog daemon
• Output format: structured Markdown wiki nodes with YAML frontmatter
• Best fit: Obsidian-friendly personal knowledge bases, research archives, and source-driven synthesis workflows
• Chinese docs: README.zh-CN.md

Idea Origin

This project is an implementation of the LLM Wiki idea popularized by Andrej Karpathy.

• Idea source: docs/karpathy_design_pattern.md
• Original concept summary: persistent wiki maintenance by an LLM instead of query-time-only RAG
• This repository adapts that pattern into a local-first, Dockerized, LM Studio-backed workflow

Highlights

• Local-first ingestion with LM Studio
• Automatic inbox watcher powered by watchdog
• OpenAI-compatible model interface
• Structured Markdown output with YAML frontmatter
• Obsidian-friendly knowledge storage under wiki_nodes/
• Source archiving after successful ingest
• Startup backlog scan plus live file watching