Most RAG setups re-derive answers from scratch; that wastes tokens and makes historical context ephemeral. LLM Wiki instead compiles sources into a persistent, incremental wiki so the LLM reasons over maintained knowledge that grows and links over time — making search, synthesis, and discovery repeatable and auditable.
What Sets It Apart
- Two-step chain-of-thought ingest: the LLM first analyzes a source (entities, contradictions, links) and then generates structured wiki pages with YAML frontmatter and source traceability, reducing hallucinations and improving maintainability.
- Graph-first retrieval: a 4-signal relevance model (direct wikilinks, source overlap, Adamic–Adar, type affinity) plus Louvain community detection surfaces clusters, surprising cross-connections, and knowledge gaps for targeted "Deep Research." Vector semantic search (LanceDB) is optional and integrates into the hybrid pipeline.
- Production-ready desktop ergonomics: cross-platform Tauri app with activity queue, Chrome web clipper, document parsers (PDF/DOCX/PPTX), async human-in-the-loop review, and a local HTTP API (127.0.0.1:19828) so agents like Claude Code can query the wiki safely.
Who It's For & Trade-offs
Great fit if you want a local, auditable knowledge base that an LLM maintains for you — researchers, analysts, and teams that need incremental synthesis and graph-driven discovery. Look elsewhere if you need a lightweight note app (this is a full ingestion + graph system), cannot configure an LLM provider or prefer a cloud-hosted SaaS. Note: optional MinerU cloud parsing uploads PDFs for complex layouts (keep built-in parser for sensitive docs), and vector search requires an embedding endpoint and DB configuration.
Where It Fits
Use LLM Wiki when persistent, provable knowledge and topological insights (communities, bridge nodes, isolated pages) matter more than ad-hoc retrieval. It complements Obsidian-like vaults by automatically generating and maintaining the pages and exposing a programmatic API for agents and automation.
