LLM-Wiki

한국어 · 中文 · 日本語 · Русский · Español · Français · Deutsch

Compile your sources into a typed wiki agents can read.

Live demo · Docs · MCP setup · Obsidian export

LLM-Wiki is a project-memory compiler. Point it at a directory containing markdown, source files, and (optionally) PDFs/Office docs/images, and it extracts a typed knowledge graph, writes a queryable wiki, and emits portable artifacts: a markdown projection, a Cognee-ready bundle, an agent harness, and an MCP server you can wire into Claude Code, Codex, or any MCP client. It is a build step for project context, not a hosted service.

How it compares

A flat comparison against the four closest open-source alternatives. No softening:

Feature	LLM-Wiki	Quartz	Logseq	Cognee	Foam
Static HTML output	yes	yes	partial (export)	no	partial (publish)
Built-in graph view	yes	yes	yes	yes (separate UI)	yes (VSCode)
Typed node schema	yes (41 types)	no	partial (tags)	yes	no
Concept extraction from sources	yes (LLM)	no	no	yes	no
Multimodal ingestion (PDF/image)	yes (via RAG-Anything)	no	partial (embeds)	yes	no
Code-graph ingestion	yes	no	no	partial	no
MCP server	yes	no	no	yes	no
Multi-project registry	yes	no	yes (graphs)	partial	no
Works without API key (OAuth)	yes	n/a	n/a	no	n/a
Multi-language i18n docs	yes	partial	yes	partial	partial
Deterministic byte-identical compile	yes	yes	n/a	no	n/a
Per-page ask widget (proposed B3)	not yet	no	no	no	no
Live edit	no	partial	yes	n/a	yes
Mobile-first reading	no	yes	yes	n/a	n/a
Real-time collaboration	no	no	yes (DB beta)	no	no

LLM-Wiki picks compile-from-source over live editing. If you want to edit notes in a UI, use Logseq or Obsidian. If you want a build tool for your knowledge graph, this is the project.

When to use this (and when not to)

Use it if:

• You want a durable, inspectable knowledge graph over a single project's text-heavy sources (docs, code, research notes).
• You want a local MCP server that answers questions grounded in your own files.
• You want to feed a clean bundle into Cognee, or a markdown projection into Obsidian, without writing the glue yourself.

Skip it if:

• You only need a vector search over a small directory — ripgrep plus an embedding library is simpler.
• You want a hosted wiki with editing UI. The static site here is read-only.
• You need accurate semantic embeddings out of the box. The default RAG-Anything embedding is deterministic (see Limitations).
• You expect a turnkey "ask anything" agent. This builds the substrate; you still wire it into your agent of choice.

Status

This is an evolving research/agent-tooling project. Known limitations:

• Compile time scales roughly linearly with corpus size. First-run compiles over large markdown trees (thousands of files) can take minutes.
• The default RAG-Anything embedding provider is deterministic. It is reproducible and dependency-free, but semantic recall is limited. Switch to ollama (e.g. qwen3-embedding:0.6b) or an OpenAI-compatible endpoint for better retrieval — see docs/integrations/rag-anything.md.
• Vision support for RAG-Anything (image content extraction) is not yet wired end-to-end. Image files are parsed structurally but not described.
• Cognee runtime cognify is best-effort: missing providers, paid API keys, or network failures are logged and skipped rather than aborting the build.
• The MCP server exposes a stable set of tools, but the underlying graph schema is still subject to additions.

Quickstart

Requires Python 3.9+. RAG-Anything needs Python 3.10+ if you enable it.

pip install llm-research-wiki

cd /path/to/my-project
llm_wiki project setup
llm_wiki project compile
llm_wiki project ask "Where is Mermaid rendering implemented?"
llm_wiki project build-site && llm_wiki project serve --port 8765

The setup wizard detects common sources (README.md, docs/, src/, data/) and writes .llm-wiki/config.json. LLM-calling features default to the codex CLI over OAuth, so no API keys are required for the common path. See docs/quickstart.md and docs/installation.md for the longer version.

Walkthrough

Each step in the Quickstart, recorded against the bundled 135-doc demo corpus (examples/demo-corpus/data/research/). Rebuild any of these GIFs with vhs docs/screencasts/<name>.tape — the tape files document what they recorded and the workspace they assume.

1. Setup — point at a research directory, get a project wiki scaffold

2. Compile + build site — deterministic, no LLM calls

3. Ask — query the compiled wiki from the CLI

What you get after compile

.llm-wiki/
  config.json
  graph.json              # typed nodes/edges
  manifest.json           # source fingerprints (used by --changed-only)
  sqlite.db               # queryable graph store
  temporal_facts.jsonl
  graphiti_episodes.jsonl
  report.md
  markdown_projection/    # human-readable wiki pages
  obsidian_vault/         # ready to drop into Obsidian
  agent_harness/          # per-agent config (Claude/Codex/Gemini/Cursor/...)
  harness_sessions/       # imported Claude/Codex session memory
  cognee_bundle/          # JSONL ready for Cognee ingest
  site/                   # static site built by build-site
  external/               # companion-tool outputs (UA, RAG-Anything)

ls .llm-wiki/ after project compile to verify what landed.

CLI overview

Daily-use commands. Run llm_wiki <subcommand> --help for full flags.

Command	What it does
llm_wiki project setup	Interactive wizard. Writes .llm-wiki/config.json. Accepts --with-understand-anything, --with-raganything, --run-cognee, etc.
llm_wiki project compile	Reads configured sources, runs companion refreshes, writes all artifacts under .llm-wiki/. Use --changed-only for incremental rebuilds.
llm_wiki project build-site	Builds the static frontend at .llm-wiki/site/.
llm_wiki project serve --port 8765	Serves the static site locally and exposes /api/ask so every detail page's inline ask widget can route questions to ask_project. On any other host (file://, GitHub Pages, S3) the widget gracefully collapses to a one-line static footer.
llm_wiki project refresh-understand-anything	Runs LLM-Wiki's managed Understand Anything refresh wrapper.
llm_wiki project refresh-raganything --parser mineru	Re-parses non-code sources (PDFs, Office, images) via RAG-Anything.
llm_wiki project ask "<question>"	Asks the configured backend (auto/raganything/cognee/wiki).
llm_wiki project mcp-config	Prints an MCP server config snippet you can paste into Claude Code, Codex, or Hermes.
llm_wiki wiki register <path> --name <alias>	Registers a project in the shared registry.
llm_wiki wiki list / llm_wiki wiki activate <name>	Lists registered projects; sets the active one.
llm_wiki ask "<question>" [--wiki <name>]	Top-level ask that resolves through the registry.

Integrations

All integrations are opt-in. None are required to use LLM-Wiki on a plain markdown/code project.

• Understand Anything — a separate project (Lum1104/Understand-Anything) that produces a code knowledge graph at .understand-anything/knowledge-graph.json. Enable with --with-understand-anything. LLM-Wiki stores a managed refresh wrapper so project compile keeps the graph current. See docs/integrations/understand-anything.md.
• RAG-Anything — multimodal ingestion (HKUDS/RAG-Anything) for PDFs, Office documents, and images via MinerU/Docling/PaddleOCR. Enable with --with-raganything. Also acts as a runtime question backend (LightRAG). Requires Python 3.10+. See docs/integrations/rag-anything.md.
• Cognee — graph+vector memory backend. Enable with --run-cognee --install-cognee. The normal compile always writes .llm-wiki/cognee_bundle/; the runtime cognify pass is best-effort and only runs when explicitly enabled.

Multi-project registry

A persistent registry at ~/.llm-wiki/registry.json lets the top-level ask CLI and the MCP server resolve project names to roots without --project on every call.

llm_wiki wiki register /path/to/my-project --name myproj
llm_wiki wiki activate myproj
llm_wiki ask "Where is the parser entry point?"

The same registry is read by the MCP server, so MCP clients can call list_projects, activate_project, and ask against any registered wiki.

Cross-vault linking (wiki:// URI scheme)

Source markdown in one registered project can reference a node in another registered project via a stable URI:

wiki://<alias>/<kind>/<slug>

Examples:

• wiki://research/concepts/rlhf — the RLHF concept in the research vault.
• wiki://other-vault/papers/arxiv-2510-12323 — a paper in other-vault.
• [See RLHF in research](wiki://research/concepts/rlhf) — works inside a Markdown link too.

At compile time these URIs become bridge nodes in the graph view (group external, violet) with a "Cross-project bridges" toggle in the toolbar so you can hide them. Unregistered aliases render as tombstones; registered-but-not-yet-built links render as placeholders.

Querying across vaults (--scope all-registered)

llm_wiki ask and the MCP ask tool accept a --scope flag:

# Default — just the active/named project.
llm_wiki ask "..."

# Fan out across every registered project; aggregate envelopes by alias.
llm_wiki ask "..." --scope all-registered

# Restrict to a hand-picked subset of registered aliases.
llm_wiki ask "..." --scope all-registered --scope-aliases research work

The aggregated JSON shape is {"scope": "all-registered", "question": ..., "by_project": {"<alias>": <envelope>, ...}}. Per-project failures are captured as {"error": "..."} entries; a single failing project never aborts the fan-out.

MCP

llm_wiki project mcp-config prints a server entry you can paste into Claude Code, Codex, or any MCP-aware client. The server exposes tools including schema, graph_summary, search_nodes, node_context, search_facts, timeline, wiki_page, raw_source, lint_report, ask, and the registry tools list_projects / register_project / activate_project / unregister_project. Tools that need a specific project resolve through the same registry as the CLI.

Authentication and LLM providers

The common path uses no API keys:

• Codex CLI (default) over OAuth. --raganything-llm-provider codex is the default; Cognee codex_cognify mode patches Cognee's LLM client to the Codex CLI.
• Claude Code CLI over OAuth. Set --raganything-llm-provider claude for RAG-Anything runtime queries. Multi-account setups use --raganything-claude-config-dir ~/.claude-personal2 (LLM-Wiki exports CLAUDE_CONFIG_DIR before each call).
• Embeddings default to a deterministic in-process provider. Switch to Ollama with --cognee-embedding-provider ollama --cognee-ollama-embedding-model qwen3-embedding:0.6b, or wire OpenAI-compatible endpoints — both documented in the integration pages.

If you set ANTHROPIC_API_KEY or OPENAI_API_KEY they will be picked up by the corresponding paths, but they are not required.

Project layout

llm_wiki/        # the package (CLI, compiler, MCP server, adapters)
docs/            # English docs + docs/i18n/ for the six other languages
ontology/        # node/edge schemas the compiler validates against
prompts/         # extraction and synthesis prompts
scripts/         # maintenance scripts
tests/           # pytest suite
evals/           # graph quality eval harnesses
data/            # example research notes used by self-dogfooding

Localized docs

한국어 · 中文 · 日本語 · Русский · Español · Français

Long-form docs are mirrored under docs/i18n/ and docs/i18n/integrations/.

License

MIT. See LICENSE.