ddsyasas/llm-wiki

First stable release of LLM Wiki — a from-scratch implementation of Andrej Karpathy's LLM Wiki pattern. Open source, local-first, bring-your-own-key, MIT.

What v1.0 ships

The three operations (Karpathy's pattern, wired end-to-end with cost previews and error recovery):

• Ingest — paste / file / URL / PDF / image → LLM compiles into cross-linked pages, refreshes the index, logs the change. Each ingest is a refactor pass, not an append.
• Query — one-shot Q&A against the wiki with cited pages. "Save as wiki page" promotes useful answers.
• Lint — local scan (broken links / orphans) + LLM pass (contradictions / gaps / stale claims / missing pages) with one-click fixes including LLM-powered ones that write the page edit for you.

Workflow surfaces:

• /sources — add via paste / drag-drop / URL, with source-detail pages showing the raw bytes + which wiki pages they contributed to
• /wiki — card grid grouped by page type, search/filter sidebar, per-page view with backlinks + source lineage + inline split-pane editor
• /graph — new in v1.0 — 3D force-directed graph of every page and every [[wikilink]]. Same engine as Obsidian's 3D Graph plugin, but colored by page type (concept / entity / comparison / overview / source) so the structure of your knowledge is visible at a glance. Click-to-focus, URL-state, bidirectional walking
• /chats — multi-turn threads saved as .md files in folders. Per-message Save-as-wiki-page + whole-chat Ingest→wiki
• /schema — split-pane editor for the CLAUDE.md contract the LLM reads on every operation
• /log — chronological timeline of every ingest / edit / lint / schema-save with clickable wikilinks
• /about /help /developers — in-app documentation reachable from the footer

Quality / safety:

• First-run wizard collects wiki topic + OpenRouter key (no silent failures on first ingest)
• Page-history backups on every edit
• Cost transparency before every LLM call, cumulative tally in Settings
• Source lineage — bidirectional traversal between wiki pages and raw inputs
• Index auto-refreshes on page edits + manual Rebuild Index button

Settings: five model slots per-operation (ingest / query / chat / lint / vision), curated dropdowns + custom-slug field, light/dark/auto theme, OpenRouter key in OS keychain when available.

Stack

TypeScript strict · Next.js 14 (App Router) · Tailwind + shadcn-style primitives · better-sqlite3 + FTS5 · OpenRouter via openai SDK · zod for LLM contracts · gray-matter · chokidar · keytar · react-force-graph-3d + three.js for the graph · vitest · pnpm workspaces

Test suite: ~140 core + 17 llm + 11 ingestion = ~168 passing tests.

Install + run

git clone https://github.com/ddsyasas/llm-wiki.git
cd llm-wiki
pnpm install
pnpm --filter @llm-wiki/web dev

Open http://localhost:3000 → the first-run wizard walks you through the rest. Full README at https://github.com/ddsyasas/llm-wiki.

Deferred to V1.x / V2

Honest about what's not in this release (tracked in docs/dev-log.md open questions):

• Per-page diff view when LLM updates a page
• Approval gate for ingest (preview changes before applying)
• Export wiki to zip
• Per-source re-ingest / delete + extracted-markdown view for binary formats
• Production build (next build) — v1.0 currently ships via pnpm dev
• CLI npm publish — pnpm pack workspace:* deps blocker
• Tauri desktop installer (V2)
• MCP server mode (V3)
• 2D toggle / search overlay / persistent camera state for the graph

Credits

Built by @ddsyasas. Pattern by Andrej Karpathy. MIT.