shannhk/llm-wikid

What it solves

LLM Wikid is designed to move beyond traditional RAG (Retrieval-Augmented Generation) for personal knowledge management. Instead of re-calculating answers from raw chunks of data every time, it creates a structured, AI-maintained wiki in Obsidian that compounds in value as more information is added and queried.

How it works

The system uses an AI agent (such as Claude Code or OpenClaw) guided by a schema in CLAUDE.md to process raw inputs into a structured knowledge base.

1. Ingest Pipeline: Raw sources (tweets, articles, PDFs, YouTube transcripts) are sorted, resolved to full content, and classified.
2. Compilation: The agent extracts key information based on the source type and creates wiki pages with cross-references ([[wikilinks]]), bias checks, and a master index.
3. Query Loop: When a user asks a question, the agent synthesizes an answer from the wiki and files that answer back into the system, allowing the wiki to grow through use.
4. Maintenance: Tools like /wiki-lint check for contradictions, broken links, and stale content, while Git ensures all changes are version-controlled.

Who it’s for

Researchers, thinkers, and power users who want a permanent, structured AI-assisted knowledge base that lives in Obsidian and avoids the ephemeral nature of RAG.

Highlights

• Anti-RAG Approach: Compiles knowledge once into structured pages rather than chunking and searching on the fly.
• Automated Ingest: Supports automated resolution of YouTube transcripts, X/Twitter threads, and web pages.
• Quality Controls: Includes mandatory "Counter-arguments" and "Data gaps" sections to prevent AI hallucinations and bias.
• Verification Gate: Pages are marked as explored: false until human-verified.
• Local Search: Integrates with qmd for hybrid BM25/vector search on-device.
• Git-Synced: Uses GitHub as the source of truth for versioning and multi-agent collaboration.