VectifyAI/OpenKB

OpenKB: Open LLM Knowledge Base

What it solves

OpenKB transforms raw, fragmented documents into a structured, interlinked wiki-style knowledge base. Unlike traditional RAG, which retrieves information from scratch for every query, OpenKB compiles knowledge into a persistent, evolving wiki where summaries, concepts, and entities are automatically synthesized and cross-referenced, allowing knowledge to compound over time.

How it works

OpenKB uses a two-layer architecture: a wiki foundation and generators.

  1. Wiki Foundation: It processes input documents (PDFs, Word, PPT, URLs, etc.) using markitdown for short files and PageIndex for long documents (PDFs $\ge$ 20 pages) to create a hierarchical tree index. An LLM then compiles these into Markdown files containing summaries, concept pages, and entity pages (people, organizations, places, products) following the Google OKF specification.
  2. Generators: These tools read from the compiled wiki to produce specific outputs, such as grounded answers via query and chat, interactive 3D knowledge graphs via visualize, and portable agent skills via the Skill Factory.

Who it’s for

It is designed for users who need to manage large volumes of complex information across various formats and others who want to create specialized, redistributable AI agent skills based on a private knowledge base.

Highlights

  • Vectorless Retrieval: Uses reasoning-based tree indexing via PageIndex instead of a traditional vector database.
  • Obsidian Compatibility: The wiki is stored as plain Markdown files with wikilinks, making it natively compatible with Obsidian.
  • Skill Factory: Distills knowledge into portable skills that can be installed by agents like Claude Code, Codex, and Gemini.
  • Multi-modal Support: Capable of retrieving and understanding figures, tables, and images.
  • Broad Format Support: Handles PDF, Word, Markdown, PowerPoint, HTML, Excel, CSV, and URLs.