wikimem

llmwiki

Build self-improving knowledge bases with LLMs.

Drop files in. Get a structured, interlinked wiki out. It improves itself while you sleep.

npx llmwiki init my-wiki

raw/                        wiki/
  2026-04-07/                 index.md ........... content catalog
    paper.pdf     ──LLM──>   sources/paper.md ... summary + citations
    podcast.mp3               entities/openai.md . people, orgs, tools
    screenshot.png            concepts/rag.md .... ideas + frameworks
    blog-url                  syntheses/ ......... cross-cutting analysis

llmwiki processes any source (text, PDF, audio, video, images, URLs), compiles it into an interlinked markdown wiki with frontmatter and [[wikilinks]], and opens directly in Obsidian.

Works with Claude, OpenAI, or Ollama (local). Your data stays on your machine.

Inspired by Andrej Karpathy's LLM Wiki pattern.

Install

npm install -g llmwiki

Requirements: Node.js >= 18 · An LLM API key (or Ollama running locally)

Quick Start

# 1. Create a vault
llmwiki init my-wiki
cd my-wiki

# 2. Ingest something
llmwiki ingest https://en.wikipedia.org/wiki/Large_language_model
llmwiki ingest ~/Documents/research-paper.pdf

# 3. Ask questions
llmwiki query "What are the key differences between RAG and compiled knowledge?"

That's it. Your wiki is now a folder of markdown files you can open in Obsidian, VS Code, or any text editor.

Why llmwiki?

The problem: You have dozens of sources — papers, podcasts, articles, screenshots, meeting recordings. They sit in folders. You forget what's in them. When you need something, you search and re-read.

RAG approach: Chunk documents, embed them, retrieve at query time. Lossy, opaque, and the "knowledge" lives in a vector database you can't read.

llmwiki approach: Compile sources into structured markdown pages with summaries, cross-references, and citations. The knowledge is readable, editable, version-controlled, and improves itself over time.

	RAG	llmwiki
Storage	Vector embeddings	Plain markdown files
Readable?	No (opaque vectors)	Yes (open in any editor)
Editable?	Rebuild index	Edit any page
Version control	Difficult	git diff
Self-improving	No	Yes (LLM Council)
Works offline	Depends	Yes (with Ollama)
Obsidian	Plugin required	Native ([[wikilinks]])

Features

• Multi-format ingestion — text, PDF, audio, video, images, URLs, Office docs
• Multi-model support — Claude, OpenAI, Ollama (local)
• Obsidian-native — [[wikilinks]], YAML frontmatter, opens directly as vault
• Semantic dedup — rejects near-duplicate sources automatically
• BM25 search — zero-dependency full-text search, no external services
• Auto-indexing — index.md + log.md maintained automatically
• Watch mode — drop files into raw/, auto-ingested
• Self-improvement — LLM Council scores your wiki and fixes issues
• External scraping — pull from RSS, GitHub, URLs on a schedule
• Health checks — find orphan pages, broken links, missing summaries
• File-back answers — query results saved as synthesis pages
• Schema co-evolution — AGENTS.md evolves with your wiki
• Domain templates — personal, research, business, codebase
• Local-first — everything is files. No database. No cloud dependency.

Architecture

┌────────────────────────────────────────────────────┐
│                    llmwiki CLI                      │
│                                                    │
│  llmwiki init         Create a new vault           │
│  llmwiki ingest       Process source → wiki pages  │
│  llmwiki query        Ask questions with citations  │
│  llmwiki lint         Health-check the wiki        │
│  llmwiki watch        Auto-ingest on file drop     │
│  llmwiki scrape       Fetch from external sources  │
│  llmwiki improve      Self-improvement cycle       │
│  llmwiki status       Vault statistics             │
├──────────────────────┬─────────────────────────────┤
│   Three Layers       │   Three Automations         │
│                      │                             │
│   raw/               │   A1: Ingest & Process      │
│   (immutable)   <────│   file/URL → markdown       │
│                      │   → place in wiki/          │
│   wiki/              │                             │
│   (LLM-owned)   <────│   A2: External Scrape       │
│                      │   RSS, GitHub, web → raw/   │
│   AGENTS.md          │                             │
│   (schema)      <────│   A3: Self-Improve          │
│                      │   LLM Council → score → fix │
├──────────────────────┴─────────────────────────────┤
│                  LLM Providers                     │
│   Claude (Anthropic) · OpenAI (GPT) · Ollama      │
├────────────────────────────────────────────────────┤
│                  Processors                        │
│   Text · PDF · Audio · Video · Image · URL · HTML  │
└────────────────────────────────────────────────────┘

Three Layers

• raw/ — Immutable source documents. Date-stamped subdirectories. Never modified by the LLM.
• wiki/ — LLM-generated markdown. Source summaries, entity pages, concept pages, synthesis pages. The LLM owns this entirely.
• AGENTS.md — Schema file. Tells the LLM how the wiki is structured, what conventions to follow, how to process sources. Co-evolved by you and the LLM.

Three Automations

• A1: Ingest & Process — Detects file type, runs the appropriate processor (Whisper for audio, ffmpeg+Whisper for video, Claude Vision for images, text extraction for PDF), asks the LLM to produce wiki pages with cross-references.
• A2: External Scrape — Fetches from RSS feeds, GitHub trending, web URLs. Deposits results in raw/ and triggers A1 automatically.
• A3: Self-Improve — LLM Council evaluates wiki quality across 5 dimensions (coverage, consistency, cross-linking, freshness, organization), proposes improvements, and applies them if below a configurable threshold.

All Commands

llmwiki init [directory]

Create a new vault with the standard directory structure.

llmwiki init my-wiki                    # Create in my-wiki/
llmwiki init .                          # Initialize current directory
llmwiki init my-wiki --template research   # Use research template
llmwiki init my-wiki --force            # Overwrite existing

Templates: personal (default), research, business, codebase

llmwiki ingest <source>

Process a file or URL into wiki pages.

llmwiki ingest paper.pdf                # PDF → extract text → wiki pages
llmwiki ingest podcast.mp3              # Audio → Whisper transcription → wiki
llmwiki ingest screenshot.png           # Image → Claude Vision description → wiki
llmwiki ingest lecture.mp4              # Video → ffmpeg → Whisper → wiki
llmwiki ingest article.md               # Markdown → wiki pages
llmwiki ingest data.json                # JSON → code block in wiki
llmwiki ingest page.html                # HTML → strip tags → wiki
llmwiki ingest report.docx              # Office → basic extraction → wiki
llmwiki ingest https://example.com/post # URL → Firecrawl/fetch → wiki
llmwiki ingest raw/2026-04-07/file.md   # Re-ingest from raw/

Each source is auto-detected by file type, copied to raw/{date}/, checked for duplicates, compiled into wiki pages by the LLM, and indexed. Use -p to pick a provider, -m for a specific model, --verbose for detailed output.

llmwiki query <question>

Ask a question and get an answer synthesized from your wiki.

llmwiki query "What are the main themes across my sources?"
llmwiki query "Compare approaches to knowledge management" --file
llmwiki query "Who is mentioned most frequently?" -p openai

Use --file to save the answer as a synthesis page in wiki/syntheses/. The query engine uses BM25 search to find relevant pages, reads the top 10, and synthesizes an answer with [[wikilink]] citations.

llmwiki lint

Health-check the wiki for structural issues.

llmwiki lint                  # Check for issues
llmwiki lint --fix            # Auto-fix where possible

Checks for:

• Orphan pages (no inbound [[wikilinks]])
• Broken wikilinks (links to non-existent pages)
• Pages missing frontmatter summaries
• Near-empty pages (< 10 words)

Reports a quality score out of 100.

llmwiki watch

Watch the raw/ directory and auto-ingest new files.

llmwiki watch                 # Watch current vault
llmwiki watch -v ./my-wiki    # Watch a specific vault

Uses chokidar for reliable cross-platform file watching. Waits for writes to stabilize before ingesting (2-second debounce). Press Ctrl+C to stop.

llmwiki scrape

Fetch content from configured external sources and deposit in raw/.

llmwiki scrape                # Run all configured sources
llmwiki scrape -s "HN Top"   # Run a specific source

Sources are configured in config.yaml:

sources:
  - name: "HN Top Stories"
    type: rss
    url: "https://hnrss.org/frontpage"

  - name: "GitHub Trending TS"
    type: github
    query: "stars:>100 created:>7d language:typescript"

  - name: "Company Blog"
    type: url
    url: "https://example.com/blog"

Supported source types: rss, github, url

llmwiki improve

Run the self-improvement cycle (Automation 3).

llmwiki improve                   # Evaluate and improve
llmwiki improve --dry-run         # Show what would change
llmwiki improve --threshold 90    # Stricter quality bar

The improvement cycle:

• Score — Evaluates 5 quality dimensions (coverage, consistency, cross-linking, freshness, organization)
• Decide — If score < threshold (default 80), improvements are needed
• Improve — Proposes actions: add cross-links, create missing pages, expand stubs, flag contradictions
• Log — Records what changed and why in log.md

llmwiki status

Show vault statistics at a glance.

llmwiki status

llmwiki vault status
────────────────────────────────────
  Pages:        42
  Words:        18,340
  Sources:      15
  Wiki links:   127
  Orphan pages: 2
  Last updated: 2026-04-07

Configuration

After llmwiki init, your vault contains a config.yaml where you set the LLM provider, external sources, self-improvement schedule, and processing options.

See docs/configuration.md for the full reference.

Environment Variables

Variable	Purpose
ANTHROPIC_API_KEY	Claude API access (default provider)
OPENAI_API_KEY	OpenAI API access
OLLAMA_BASE_URL	Ollama server URL (default: http://localhost:11434)
FIRECRAWL_API_KEY	Enhanced URL-to-markdown (optional, falls back to fetch)
DEEPGRAM_API_KEY	Audio transcription API (optional, falls back to Whisper)

Multi-Model Support

llmwiki works with any major LLM provider. Choose at init time or per-command.

Provider	Flag	Default Model	Env Variable
Claude	-p claude	claude-sonnet-4-20250514	ANTHROPIC_API_KEY
OpenAI	-p openai	gpt-4o	OPENAI_API_KEY
Ollama	-p ollama	llama3.2	OLLAMA_BASE_URL

# Use Claude (default)
llmwiki ingest paper.pdf

# Use OpenAI
llmwiki ingest paper.pdf -p openai -m gpt-4o-mini

# Use Ollama (fully local, no API keys)
llmwiki ingest paper.pdf -p ollama -m llama3.2

Multi-Format Support

Format	Extensions	Processor	Requirements
Text	.md, .txt, .csv	Direct read	None
PDF	.pdf	Built-in text extraction	None
Audio	.mp3, .wav, .m4a, .ogg, .flac, .aac	Whisper / Deepgram	whisper CLI or DEEPGRAM_API_KEY
Video	.mp4, .mov, .avi, .mkv, .webm	ffmpeg + Whisper	ffmpeg + whisper
Image	.jpg, .png, .gif, .webp	Claude Vision	ANTHROPIC_API_KEY
HTML	.html, .htm	Tag stripping	None
JSON	.json	Code block wrapping	None
Office	.docx, .pptx, .xlsx	Basic extraction	None (enhanced coming)
URL	https://...	Firecrawl / fetch	Optional FIRECRAWL_API_KEY

When a processor's requirements are not met (e.g., Whisper not installed for audio), llmwiki creates a reference page noting the source file and suggests installing the missing tool. The raw file is always preserved.

Obsidian Integration

llmwiki vaults are Obsidian vaults. Open any llmwiki directory in Obsidian and you get:

• Graph view showing all pages and their [[wikilinks]]
• YAML frontmatter rendered as page metadata
• Backlinks panel showing what links to each page
• Search across all wiki content
• Tag view from frontmatter tags: arrays

No plugins required. No configuration. Just Open folder as vault in Obsidian.

# Every wiki page has frontmatter like this:
---
title: "Attention Is All You Need"
type: source
created: "2026-04-07"
updated: "2026-04-07"
tags: [transformers, attention, nlp]
sources: ["raw/2026-04-07/attention-paper.pdf"]
summary: "Foundational transformer architecture paper introducing self-attention"
---

Vault Structure

my-wiki/
├── AGENTS.md             # Schema — wiki structure + conventions
├── config.yaml           # Configuration — provider, sources, schedules
├── .gitignore
├── raw/                  # Immutable source archive
│   ├── 2026-04-07/
│   │   ├── paper.pdf
│   │   ├── podcast.mp3
│   │   └── blog-post.md
│   └── 2026-04-08/
│       └── meeting-notes.md
└── wiki/                 # LLM-generated knowledge base
    ├── index.md          # Auto-maintained content catalog
    ├── log.md            # Chronological operation record
    ├── sources/          # One summary per ingested source
    ├── entities/         # People, organizations, tools
    ├── concepts/         # Ideas, frameworks, patterns
    └── syntheses/        # Cross-cutting analyses, query results

Tests

cd /path/to/llmwiki && pnpm test

Contributing

See CONTRIBUTING.md for guidelines.

See Also

• agentgrid — Manage grids of AI coding agents in tmux
• agentdial — Universal agent identity protocol across 8 messaging channels

Credits

Inspired by Andrej Karpathy's LLM Wiki pattern — the idea that LLMs should compile knowledge into structured, interlinked wikis rather than just answering questions from raw chunks.

License

MIT — see LICENSE.