Labhund / llm-wiki

What is llm-wiki?

llm-wiki is an agent-first knowledge base that turns documents into a plain markdown wiki with wikilinks, beating RAG by compounding insights over time. Ingest PDFs or papers via CLI, and it extracts concepts into linked pages; a Python daemon with background agents (auditor, librarian, adversary) keeps everything indexed, cited, and verified against sources. Agents navigate via MCP server tools like wiki_search, wiki_read, and wiki_query, mimicking Wikipedia-style traversal without token waste.

Why is it gaining traction?

Unlike RAG's per-query resets or raw markdown dumps, this builds a living graph: ingest once, query cited answers that improve with use, thanks to synthesis caching and git-committed agent writes. Background quality agents surface issues inline during reads, and Obsidian compatibility means humans edit the same files. MCP integration hooks into Hermes or Claude Code for seamless agent navigation in LLM wiki english or multilingual setups.

Who should use this?

AI researchers ingesting llm wikipedia dataset or papers for agent swarms, needing persistent knowledge beyond session RAG. Devs building agent-first orgs on GitHub, handling law docs or wikidata llm tasks with background maintenance. Teams wanting markdown vaults for navigation, where agents query "boltz" and get cited gaps like "Missing: boltz-2 details."

Verdict

Try it for agent prototypes—CLI ingest/query shines, docs are solid with philosophy deep dives—but skip production until stable (18 stars, 1.0% credibility). Active dev warns of breaks; cap LLM costs via config first.

(198 words)