benmaster82/Kwipu
Ask questions across your Markdown notes using a fully local Graph RAG engine. Built for Obsidian vaults, works with any folder of Markdown files. Extracts entity-relation triples from wikilinks & YAML frontmatter, retrieves answers via hybrid search (vector + BM25 + temporal). Multilingual. No cloud. Runs on Ollama.
What it solves
Kwipu allows users to turn a folder of documents (such as Obsidian vaults or general notes) into a searchable, interactive knowledge graph. It solves the problem of querying large sets of personal notes by providing grounded answers with citations, rather than just simple keyword searches.
How it works
Kwipu processes documents in .md, .txt, .pdf, and .docx formats. It uses an LLM to extract semantic relations and identifies structural relations via wikilinks and YAML frontmatter to build a property graph. For retrieval, it combines vector similarity, BM25, and temporal metadata. Users can interact with this data through a command-line interface, a 3D web visualization, or an MCP (Model Context Protocol) client. The system can run fully locally using Ollama or connect to cloud-based models.
Who it’s for
It is designed for individuals who manage large knowledge bases or "second brains" (like Obsidian users) who want to query their notes using AI and visualize the connections between their entities and documents.
Highlights
- Multi-interface access: Query via terminal, a 3D web UI, or an MCP server.
- Hybrid retrieval: Combines vector search, BM25, and temporal metadata for higher accuracy.
- Local-first option: Supports fully local execution via Ollama for privacy.
- Automatic updates: Watches source folders and incrementally updates the graph index.
- Grounded answers: Provides responses based on source chunks with clear citations.
Related
- Project
- Project
- Project
- Project