LLM Wiki is a cross-platform desktop application that turns a pile of personal documents into a structured, interlinked wiki which a language model writes and keeps current. It is aimed at people who keep a large private library — researchers, analysts, students, engineers sitting on a folder of specifications — and who have grown tired of question-answering setups that start from zero on every prompt. The project is written mainly in TypeScript, and it has collected roughly 18,700 stars and 2,100 forks since it first appeared in April 2026.
What it does
You point the app at your files and it produces wiki pages, not just answers. The result is a persistent knowledge base you can browse, rather than a black box you have to interrogate.
The feature list covers the parts of that job most people hit first:
- Document parsing across PDF, Office formats, EPUB and MOBI, Org mode, images, media, web clips and batches of URLs, with built-in, cloud or local MinerU handling for PDFs.
- Multimodal image ingestion that pulls images embedded in PDFs, captions them with a vision model, and surfaces them in image-aware search with a lightbox preview and a jump back to the source.
- A knowledge graph that exposes connections between pages, with automatic clustering of related material.
- A Read Sources Only mode that answers strictly from the material you imported.
- Project export and import, so a complete archive moves between machines, plus a rebuild of the wiki index from pages you already have.
The channel's video also shows it plugging into Claude Code as a skill, so the same knowledge base is reachable from a coding session.
How it works
The core idea is incremental rather than retrieval-on-demand. Conventional retrieval-augmented generation fetches chunks and composes an answer from scratch each time you ask something; LLM Wiki instead spends that effort once, at ingest, and keeps the output.
Ingest runs as a two-step chain of thought: the model analyses a document first, then generates wiki pages from that analysis, with source traceability on what it writes and an incremental cache so repeat work is not redone. Pages accumulate and link to each other, which is what the project means when it says knowledge compounds instead of resetting.
Structure comes from the graph layer. Relevance between pages is scored from four signals — direct links, overlap in source documents, Adamic-Adar, and type affinity — and Louvain community detection groups the result into clusters with a cohesion score, so topics emerge from the material rather than from folders you maintained by hand.
Model configuration is per project. Chat and Ingest can be routed to different models independently, and custom providers, headers and streaming output are all configurable, which matters because ingest is the expensive half of the workload.
Getting started
It is a desktop application, so the entry point is installing it rather than standing up a server; the README carries its own installation and tech-stack sections. Plan on two setup decisions before your first real import. The first is which models to point at Chat and at Ingest, including whether you route them to different providers. The second is how PDFs get processed — built-in, cloud, or a local MinerU instance — since that choice drives both quality and where your documents travel.
After that, start with one project and a bounded set of sources rather than your whole archive. Because ingest is cached and incremental, adding material later is cheap; a first pass over thousands of files is not. If you want to move the project to another machine afterwards, the archive export and import path is built in.
When to use it / when not
This fits a corpus you return to repeatedly: a research library, a standards collection, product documentation, a long-running case file. The payoff is that the model's reading is preserved as browsable pages with links back to sources, so you can audit an answer and find neighbouring material you did not know to ask for.
It fits less well when your documents are throwaway or change constantly, or when you only need one answer from one file — a plain retrieval tool will get there with less setup and less token spend. It is also a single-user desktop app by design, so a team looking for a shared, always-on service will be building that around it. And because ingest does real model work over every document, cost scales with the size of the library, not with how many questions you ask.
Anyone whose real complaint about AI document tools is that they forget everything between questions should look at this seriously. LLM Wiki takes the unglamorous position that reading should happen once and be written down, and it backs that up with source-linked pages, a graph that finds links you would not, and captioning that rescues the figures buried inside PDFs. The trade is an ingest bill up front and a desktop app rather than a service. If your documents are worth reading twice, that is a good trade.