LLM Wiki extracts key concepts from your PDFs, docs, and articles, weaving them into an interconnected, queryable wiki — with zero manual writing.
Built around the idea that consuming papers is only half the work — structuring what you learn is the other half.
Accepts PDF, TXT, HTML, DOCX, PPTX, and XLSX. Large PDFs are intelligently chunked so nothing gets dropped.
Every concept page is rich with [[bracket]] links to related concepts, forming a genuine knowledge graph.
Add a new paper and run /sync-wiki. Only changed or new concepts are updated — existing pages are never overwritten needlessly.
Run /audit-wiki to surface orphan pages, missing links, contradictions between sources, and stale claims.
A Streamlit chat interface queries only your wiki — no hallucinated outside knowledge. Answers cite their source page (e.g. source: bert.md). Unknown topics return an explicit "No information found" with remediation steps.
Every concept traces back to exact papers and pages. Contradictions across sources are captured as features, not hidden.
Compile, sync, audit, and reset work in both Claude Code CLI and GitHub Copilot Chat. The interactive query UI is available for both.
Your wiki lives as plain Markdown files in a git repo. Version control, diffs, and collaboration come for free.
Pair with the Obsidian Web Clipper browser extension to capture articles and blog posts directly into /raw.
Drop sources into /raw — then run the full pipeline in one command, or step through it yourself.
First run — batch-processes all sources in /raw, extracts concepts, and creates structured entity pages.
Added a new paper? Sync detects what changed and incrementally updates only the relevant pages.
Once you've grown to ~20+ pages, audit checks for orphans, broken links, contradictions, and stale claims.
/reset-wiki reset before running the pipeline to wipe them and start with your own papers.
Without the reset argument, /reset-wiki only reports state — it never deletes anything.
Every generated page follows a consistent template — definition, explanation, related concepts, and sources.
[[bracket]] links create a genuine knowledge graph you can visualise in Obsidian.All commands are available flat — no namespaces. Commands work in both Claude Code CLI and GitHub Copilot Chat (Copilot uses slightly different names). The query UI is Claude Code only.
One command to run the full pipeline — detects wiki state, compiles or syncs sources using parallel agents, audits quality with parallel agents, then syncs all doc surfaces.
Batch-processes every file in /raw, creates entity pages, and builds the index from scratch. Blocks if wiki already has content — run /reset-wiki first.
Detects files in /raw not yet logged in wiki/log.md and incrementally updates only the affected pages — existing content is never needlessly overwritten.
Scans for quality issues: orphan pages, missing linked concepts, cross-source contradictions, and stale claims. Run every ~20–30 new pages or before sharing.
Smart state guard: if wiki is empty it scaffolds the index & log; if content exists it summarises and offers options. Pass the reset argument to force-wipe all entity pages.
Checks that streamlit and litellm are installed, reports available wiki pages, then starts the Streamlit query interface at localhost:8501.
Checks whether the Streamlit process is running, kills it if active, and confirms it is stopped. Companion to /launch-wiki-ui.
Detects changed files via git diff, maps them to affected doc surfaces (README, GitHub Pages, copilot-instructions), and applies targeted edits — never rewrites entire files.
Runs the full pytest suite, verifies every skill is registered in both Claude Code and Copilot, checks wiki health and unsynced sources, inspects git status, and saves a dated report to reports/.
Detects which generator scripts or skill files changed via git, rebuilds only stale assets (terminal/UI demo videos + PPTX deck), and verifies both videos are embedded in the deck.
Run /launch-wiki-ui to open a guided 3-screen Streamlit interface. Configure your model and API key, confirm which wiki pages are loaded, then chat — answers are grounded strictly in your wiki.
raw/, then run /sync-wiki.
Clone the repo. For the query UI, have an API key ready from any major LLM provider (Anthropic, OpenAI, Google, Mistral, Cohere, etc.). No other setup required.
git clone https://github.com/dev-enthusiast-84/llm-wiki.git
Starting fresh? The repo includes example wiki pages.
Run /reset-wiki reset to wipe them before running the pipeline with your own papers.
/rawPDF, DOCX, TXT, HTML, PPTX, or XLSX — anything goes. The Attention paper, BERT, GPT-3, and Scaling Laws papers are included as examples.
Recommended: run /orchestrate-wiki — detects state, compiles or syncs, audits, and syncs docs in one go.
Or step by step: /compile-papers first run → /sync-wiki after each new paper → /audit-wiki at 20+ pages.
/launch-wiki-uiA 3-screen flow: enter your model ID and API key → confirm loaded wiki pages → chat. Answers are grounded strictly in your wiki and cite the source page. Topics not yet covered return an explicit "No information found" with instructions on which file to add.
LLM Wiki is open source, MIT licensed, and ready to use today.