atomicstrata / llm-wiki-compiler

项目摘要

llmwiki is a TypeScript CLI tool that compiles raw sources (papers, notes, PDFs, transcripts, web pages) into an interlinked, citation-traceable markdown wiki designed for both LLM agents and humans. It implements Karpathy's LLM Wiki pattern, shifting knowledge work from query-time retrieval to compile-time compilation, with durable pages that accumulate structure, provenance, and review state. Extensible lifecycle profiles (e.g., AutoSci, Newsroom) enable domain-specific typed records and workflows without forking the compiler.

为什么值得关注

It is one of the first production-grade implementations of Karpathy's LLM Wiki pattern, offering a compelling alternative to traditional RAG by compiling knowledge once into durable, provenance-backed pages. With 2,100+ stars in roughly six months, 13 releases, CI/tests, an npm package, signed template distributions, and a polished local viewer with theming, it shows unusually strong engineering discipline for a young project in the emerging context-engineering space.

优势

• Novel compile-once knowledge approach with full citation traceability and provenance back to source passages, addressing hallucination and trust issues in RAG pipelines
• Strong engineering hygiene: CI, tests, 13 releases in six months, npm distribution, docs site, and 15 contributors
• Extensible lifecycle profile system (typed records, relationships, gates, workflows) plus signed template catalogs keeps domain logic out of the compiler core
• Dual-purpose output: agents can browse, query, lint, and export the wiki while humans get a themed local viewer and Obsidian-compatible markdown
• Handles heterogeneous inputs including PDFs, images, transcripts, and web pages with recursive folders and path exclusions
• Embedding index storage enables retrieval metadata to accumulate alongside compiled pages

局限性

• Young project (about six months old) with an ecosystem of profiles and templates still maturing
• Compilation quality and cost depend on the underlying LLM, so large corpora may be expensive to compile and re-compile
• No Docker packaging, which may complicate deployment in some environments
• Profile and template concepts add conceptual overhead compared to simpler note-taking or RAG tools

使用场景

• Compiling research papers, lab notes, and transcripts into a browsable scientific wiki that AI agents can query with citations
• Building domain-specific knowledge bases (newsrooms, legal archives, internal docs) with typed records and lifecycle review gates
• Giving coding or research agents a durable, pre-compiled knowledge context instead of re-retrieving from raw files at query time
• Producing Obsidian-compatible interlinked markdown wikis from messy heterogeneous source folders