OpenKB is a command-line system that compiles raw documents into a structured, interlinked wiki and then answers questions from it. Rather than retrieving passages afresh on each query, it writes summary pages, concept pages, and entity pages for people, organisations, places, and products, keeps them cross-linked, and updates them as new sources arrive, so a single added source can touch ten to fifteen wiki pages.
Handling depends on document length. Short documents are converted to Markdown and read in full by the model; long PDFs are turned into a hierarchical tree index by PageIndex, and the model reads the tree instead of the whole text.
| Short documents | Long documents (PDF at least 20 pages) | |
|---|---|---|
| Convert | markitdown to Markdown | PageIndex tree index and summaries |
| Images | extracted inline with pymupdf | extracted by PageIndex |
| Model reads | full text | document trees |
| Result | summary and concepts | summary and concepts |
The wiki is a directory of plain Markdown files with wikilinks, so it opens in Obsidian for graph browsing, and a wiki/AGENTS.md file defines the structure and conventions the model follows when maintaining it. Model access goes through LiteLLM in provider/model form, covering OpenAI, Claude, Gemini and other providers, along with OAuth device-flow subscriptions that need no API key. PageIndex runs locally by default, with an optional cloud service for OCR on scanned PDFs and faster indexing of large documents.
On top of the wiki sit generators: one-off queries with cited answers, multi-turn chat sessions that persist and resume, an interactive knowledge graph, single-file HTML slide decks, and a skill factory that distils a portable agent skill other agents can install. A bundled web interface served by the REST API lets you browse the knowledge base, upload and compile documents, and stream queries in a browser.
Features
- Input formats: PDF, Word, Markdown, PowerPoint, HTML, Excel, CSV, text, and URLs, with content type auto-detected for URLs
- Long-document handling: PageIndex tree indexing gives vectorless, context-aware retrieval without a vector database
- Multi-modality: figures, tables, and images are retrieved and interpreted alongside text
- Compiled wiki: summaries, concept pages, entity pages, and cross-links stay in sync as documents are added, recompiled, or removed
- Query and chat: cited one-off answers or multi-turn sessions with resume, list, and delete, plus slash commands inside a chat
- Skill factory: distils a redistributable agent skill from the wiki, with validation, trigger evaluation, version history, and rollback
- Decks and graphs: single-file HTML slide decks with an optional critique pass, and a self-contained 3D, mind-map, and radial knowledge graph
- Knowledge Workbench: a bundled browser interface for browsing, uploading, compiling, and streaming queries, with optional bearer-token auth
- Open Knowledge Format: wiki pages follow the Google OKF specification for knowledge sharing
- Maintenance commands: watch a folder for new files, list documents and concepts, show statistics, and run structural and knowledge health checks
