Turn Your Markdown Pile Into a Queryable Property Graph

What it does Kwipu ingests .md, .txt, .pdf, and .docx files from any folder, including Obsidian vaults, and builds a property graph that mixes LLM-extracted semantic relations with structural ones pulled from wikilinks and YAML frontmatter. You query it through a terminal, a 3D web interface, or an MCP client, and it grounds answers in source chunks with citations. It also watches your knowledge folder and updates its persisted storage as files change.

The interesting bit The retrieval stack is deliberately over-engineered in the best way: vector similarity, BM25, temporal metadata, and optional synonym expansion all feed into the same graph. It can reconstruct project timelines and compare roles across meeting notes by leveraging that temporal layer. The 3D visualization lets you browse entities and documents as a literal web of relations rather than a flat list of search results.

Key highlights

• Hybrid search combining vector similarity, BM25, temporal metadata, and an optional LLM synonym retriever.
• Property graph built from both semantic LLM extraction and structural note features like wikilinks and YAML frontmatter.
• Multilingual pattern support for English, Italian, French, German, Spanish, and Portuguese.
• Three interfaces: interactive terminal, 3D web UI, and MCP server.
• Folder watcher that safely rebuilds persisted graph storage as documents are added, edited, or deleted.

Caveats

• The default LLM is gpt-oss:20b-cloud, which may send your document chunks and questions to a remote provider; fully local operation requires explicitly switching to a local Ollama model.
• The web interface requires a Node.js frontend build in addition to the Python runtime.
• Remote Ollama endpoints require HTTPS unless you explicitly enable insecure HTTP for a trusted network.

Verdict Worth a look if you maintain a large personal knowledge base in Markdown or Obsidian and want hybrid graph retrieval that can run entirely on your own hardware. Skip it if you need a managed cloud service or would rather not babysit an Ollama endpoint and a Python environment.